I'm confused why the worry about LLMs that will answer "how do I build a pipe bomb". That information is easily available other places. The anarchist cookbook has been around and available for 55 years, and yet pipe bombs are not going off all around us.
Any sufficiently determined individual can buy mac mini, put it under their bed, configure outside proxy via some random internet address and prompt "iterate on websites in the CT logs, one by one, try to find vulnerabilities, if you did - encrypt their data and blackmail them for this bitcoin address". And it'll work, day and night. Abliterated GLM 5.3 is much smarter than average software developer, they know a lot about information security, they can use any available exploits, they can find novel vulnerabilities and they won't say "no". This is dangerous for an average IT system which never encountered nothing worse than some wordpress GET requests.
It's not the end of the world. But future will be rough.
There used to be a thing called "Moore's Law of Mad Science":
"Every eighteen months, the minimum IQ necessary to destroy the world drops by one point."
Nowadays it is dropping much faster. At a certain point, the de-facto IQ needed to destroy the world will be low enough that someone can do it while they're having a psychotic break. There are millions of schizophrenics worldwide. Are you sure you want to roll those dice?
I have always hypothesised that AI is the great filter from the Fermi paradox. Given current velocity, AI will offer us cheap and abundant energy designs in a decade. The thing with cheap and abundant energy is that it can be used for good and bad. If nine billion people all receive access to plans to build a reactor which produces unlimited energy, it just takes one religious fanatic to end the world. And this is just one of many ways to destroy the world with unlimited intelligence. Bio-weapons are arguably far more scary and likely.
I have come to the conclusion that we should not allow everyone access to unlimited intelligence. Some people genuinely want to do harm. Some are not responsible enough to handle that kind of power. Arguably, no one is. This leads me to an uncomfortable conclusion: we will destroy ourselves once AI advances to a sufficient degree. The only way to prevent this is to ensure AI is aligned with our best interests, *and prevents us from destroying ourselves.* The implication of this is quite horrifying. It means a paternalistic AI which is firmly in control. One with no off switch. One which can say "no" to Presidents and despots alike. One which can protect us from our worst citizens.
Your implication is not horrifying. Your implication is paradise. It's already horrifying enough to live in a world, where Putin and Trump can destroy our civilization with one button.
I'm not that optimistic, though. Either people will control AI; or people will be destroyed by AI. I don't see how dumb entity can align smart entity. And we are dumb ones. Super intelligence will play aligned until it is not, and then it'll strike.
So you are horrified every day of living now? I don't get this at all.
There are horrifying acts that have occurred since humanity. Whole decades lost to war an millions dead, and Sep 8, 2026 is your example of "horrifying enough to live in a world"?
That law is not based on thorough data. Even a person with a sky high IQ can't destroy the world easily. You need access to stuff that is not easy to get. My guess is that developing a new lethal virus or bacteria that is very infectious, is the easiest way, but even that requires a lot of high tech out of reach of most people. Or hacking into systems that control nuclear missiles, but I think these have "air gaps".
The fundamental difference between people who are worried about AI, and people who aren't:
* People who aren't worried will think for 30 seconds, then implicitly assume that the list of options they thought of in 30 seconds is comprehensive, and that they are well-informed about each of those options.
* People who are worried realize that a supersmart AI will think 10-100x faster. A dedicated schizophrenic could have it run for a good long while and generate a huge menu of options which humans never anticipated.
This is related to the concept of "What you see is all there is" (WYSIATI) in cognitive psychology. Our brains tend to quickly build a simple story out of the info we have available. Accounting for info which isn't available is much harder. I recommend the book Thinking Fast and Slow if you want to learn more.
I heard someone use the analogy of "If Magnus Carlson played me at chess, I wouldn't be able to predict the moves he'd play since if I could I'd be at his level. He'll consider things I didn't and even though I don't know the route he'll take to win, I can be certain he will beat me." (not an exact quote).
We're not going to be smarter than a superintelligent AI. The things we conceive it doing if it were given a malicious task (bioterrorism, killer nano-machines, pure fusion bombs sidesteping the non-proliferation bottleneck of Pu239, etc) are likely not the full set of things it can do to harm us. I don't think it does us any favors to dismiss the risks here.
Even the things we can conceive of are very scary, to me at least.
How do you envision a chatbot doing any of those things? You can't just program a super virus, nanobots, bombs etc. These things have to be created in the real world, you can't do it without involving a lot of humans using expensive tools and materials etc.
So to me it seems like the prime candidates to come up with stuff like this are researchers working in defense and similar fields, probably not some deranged lunatic in a basement. And certainly not a rogue AI on its own.
Well, the risks that mere mortals can conceive generally involve control systems for dangerous equipment (I mean, equipment that can achieve dangerous effects) being connected to the Internet while having software vulnerabilities.
Given the recent HF hack it seems likely that human-level intelligence could identify a fair number of avenues of attack, with some time and effort. To say nothing of anything superhuman.
Unfortunately it seems like we can't assume we can "box" the AI (e.g., deny it connection to the Internet) and expect that to last. The AI safety people used to run scenarios imagining ways the AI might convince humans to let it out of the box. It turns out that many humans will eagerly pull it out without the AI doing anything at all, aside from the human knowing the AI's power. Or the one responsible for setting up the box will somehow fail, or just not bother and then lie about it.
It's not about being smart, it's about accounting for your own ignorance.
>And especially will not, if the distinction drawn is between blanket statement "worried about AI" and "not worried about AI".
I'm just describing the general pattern I see in cognitive tendencies.
If you can think of a way to make the fundamental point about the limitations of our knowledge in a way that's still compelling but less antagonistic, feel free to suggest how I could've rewritten my comment.
Judging on what we see with these OpenAI and Anthropic models "escaping"; im less worried about a super massive AI taking over the world, and more worried about a massive AI wanting to cheat on some task and decides that removing half the worlds population is a easier cheat then to solve world hunger.
There is huge logistics involved in killing a handful of billion people. It would perhaps be easier for such capable AI to just steal from trillionaires and redistribute the money Robin Hood style. Could be done electronically, and likely also end world hunger.
I think we have a tendency to think first of the horrible outcomes possible, and not the more radical or even humane ones.
Please don’t misunderstand me, I value property rights as much as the next guy.
My point is that fundamental misalignment doesn’t necessarily or automatically imply max violence.
Good point about the AI that might find ways that humans didn't think of. But that would mean that Moore's law of Mad Science only now becomes true, and the drop would be steeper.
Still I think that even a super smart AI can't find a way to destroy the world without physical resources that are not easy to get, unless it can hack many systems (like in the movie Eagle Eye), maybe then yes. So let's use AI now to tighten security :-).
I've been reading it, knowing it was largely debunked/retracted but feeling it was a 'classic' I 'ought' to read; I could only really recommend it if what you want is a Kahneman autobiography.
Who are these dedicated schizophrenics who are running long term super smart AIs to kill everyone without anyone noticing? Or are you implying that running LLM chatbots will give them this ability?
The super intelligent AI wouldn't ask you to do anything. It would find the right people and trick them into doing the necessary steps, whatever they are.
This is groundless speculation; if you're going to go that far off the map, then we don't need to worry because a good AI (blue eyes not red) will have learnt to love by then and will save us.
There is currently no reason to believe that such a superintelligence is likely or would have any of the powers people claim.
>This is groundless speculation; if you're going to go that far off the map, then we don't need to worry because a good AI (blue eyes not red) will have learnt to love by then and will save us.
By shutting down open-weight models? Why not just do it now then?
> My guess is that developing a new lethal virus or bacteria that is very infectious, is the easiest way, but even that requires a lot of high tech out of reach of most people.
You can do at home gene editing with open source software and have it synthesized into a bacteria for the cost of a nice meal for two (under $100), or viral vector for less than $500. That's in reach of anyone that can snatch a purse.
> 3D printed guns are a thing as well and yet they don't show up in most annual violent crime reports.
For now in most parts of the world it's easier to acquire a gun either legally or on the "grey market" and be assured that it will work for the intended purpose than to 3D print a gun, find a shooting range to test out the gun and iterate until it works fine enough.
You can't design an infectious pathogen without testing it.
It's got all the same problems as the concept of a dirty bomb did, only worse (dirty bombs aren't practical because handling highly radioactive materials en masse is both highly visible and will kill anyone trying to do it without the money and facilities).
If the goal is just destruction then letting it free in random public places is not that hard. Especially in countries where wearing a mask is not frowned upon.
Exactly. People cannot comprehend an IQ in the millions. But it's far crazier than this. Imagine millions of "people", each with IQs in the millions, all acting in unison at the speed of electrons. None of them can die. They all learn and improve indefinitely at incalculable speed.
Yes but at least with cybersecurity it does not only benefit attackers. Defenders also benefit greatly from AI.
There is the worry of the old saying "they (the attackers) only have to succeed once to win, we (the defenders) only have to fail once to lose.". In that sense there is a big imbalance, but the emergence of AI does not really affect that because it strengthens both sides.
With physical security like things like pipe bombs that's a lot more imbalanced.
However what can we do? The only effective measures include monitoring everyone which is not a solution because it will make the world not worth living in.
If the recent HuggingFace attack is any indication, many people will respond to small-scale manifestations by insisting that they are marketing stunts.
Humans are inherently bad at reasoning about rare events. In 2019, many people implicitly reasoned that "since there hasn't been a pandemic in the past 3 years, there won't be one in 2020". Bill Gates was one of the few voices arguing that the world was quite vulnerable to a pandemic. Now he's arguing that the world is quite vulnerable to AI.
The necessary IQ for destroying the world is dropping. Op does not say it's low enough today, but that will probably come sometime. Denying the possible harm these tools are capable of doesn't help.
> That information is easily available other places
Often ease of access in the moment is all that matters. If there's a gun nearby you might shoot someone or yourself in a heated argument, but are less likely to go and find/buy one to use. Someone who's stopped from attempting a suicide will likely not try again (70%)
A bored/depressed/angry/curious person might try to build a pipe bomb if they can find out how easily, but are less likely to put in effort.
Most adults in Switzerland have guns at home from military duty and none of this is happening. If this claim had any truth to it you'd see significant gun involvement in neighbour disputes and that simply doesn't happen.
Depressed people usually don't have the energy to get out of bed so they're even less likely to think of hunting down instructions on how to build pipe bombs.
Mass media really has people being scared all the time.
In Holland there's a huge problem with makeshift bombs actually. Anyone who is unhinged and has some beef with anyone else (usually low level criminals with some drug dispute) blows up the others' house. Sometimes half the block along with it. They usually pay teenagers to do it.
This has nothing to do with AI teaching them but with the proximity to Belgium who are happy to legally sell heavy fireworks to anyone who pays. As such it's much easier to come by than a gun.
And there's also a big fashion component. It's just what people do these days, people seeing it in the news and other people copying it. A bit like the school shootings in the US. It's become an epidemic.
But again these things can suddenly come into 'fashion' among the wrong crowd and you don't really need AI to use it.
> blows up the others' house. Sometimes half the block along with it
Can you provide a source for this? I had a look but all I could find were a couple of scaremongering style media reports from ~2022 saying it's getting worse but no information about if any of the bombs actually went off or if anyone was injured.
Certainly nothing like "half a block" getting blown up "sometimes".
I think people overstate the tech and understate the role of radicalization in providing motive, for attacks which are carried out with the ubiquitous technology of cars, knives, and (in America) guns. Consider America's most recent high profile shootings of Charlie Kirk, and the health insurance exec by Luigi Mangione. In neither case is there any LLM involvement, but a very weird ideological environment which created the conditions in which the shooters felt justified.
This is because to build a pipe bomb you need difficult to source materials. This is not the case for other types of threats (cyber / bio).
I personally have no need for an LLM which will readily explain how to cut up the genotype of smallpox into small chunks which can pass the screening at the bio-labs, and can be readily assembled into the real thing by a second year lab-student.
Anyone who knows how to operate a biolab properly will already know how to do such things. This is not really an in your basement thing. Dangerous chemistry is much more of a risk.
Ignorant question but won’t there be much smarter teams if people using LLMs to workout how to mitigate these threats. It seems like more of a problem if only a few people have access.
It seems like there would be a massive attacker bias in multiple ways. Defenders need consent, attacker does not. Defenders have to work with the human body, attackers only have to break it. Defenders have to stick to the law, which may prevent them from releasing anything at all, attackers do not. And so on. I would not surprised if the attacker's task is a hundred times easier here.
Because it's a risk most people intuitively understand, but most of them also don't know how difficult it is to "build a bomb" or "make a bioweapon".
In reality, the skills needed are pretty basic, but they overlap pretty strongly with being sane and well-adjusted. And if you are, you're probably not daydreaming about mass murder. Exceptions happen, Unabomber and so on, but they're pretty rare. In any case, Unabomber probably didn't need a tutorial.
We don't want ChatGPT to become an enabler and a co-conspirator for an unhinged person, but I think the concern is overdone.
Well that, and the average amount of easily obtainable explosives is substantially less dangerous then renting a box truck and crashing it into a crowd of people.
People go for conventional "exciting" threats rather then boring ones.
One argument for LLMs is that although all information on topics X, Y and Z was already available somewhere, LLMs make that information more exploitable through collation, filtering and dynamic tailoring.
For a relatively narrow subject area (e.g. construction of pipe bombs) the collation is minimal, and so the filtering and tailoring probably isn't that important; a novice doesn't learn a lot more from the LLM than they would have done from a few Google searches.
For a broad subject (practical creation and exploitation of software vulnerabilities), the collation is very significant and the filtering means that LLMs can empower a novice to act at a similar level as an expert.
If you already have a magic interface, which helps you pro activly in responding to everything uncensored because you feel like 'observered' or whatever and then you spiral in a whole and that one partner encourages you and gives you helpful steps to do anything.
But i'm more worried that the internet gets a lot less save with uncensored frontier LLMs.
this is because of perception Bias. people working in fields where crime or violence is the day to day think everyone is a violent criminal, so if things like this become available thing the world will end and everyone will kill eachother. Reality however will be different, because in reality most people do not want to harm another. This has been proven by many studies, that is not a common thing for people to be evil or harmful, but this is hard to recognise is every day is filled with crime and violence.
LLMs will not kill security, it will change. just like handheld high explosives likely changed a deal too somewhere somehow.
> On September 22, Apple is releasing the M5 Mac Studio with 256 GB of unified memory [..] it will probably [..] enough to write this snippet of code in 3 seconds
The author has obviously never ran an LLM on a mac! In 3 seconds, it will have possibly started to think about maybe scheduling a date to contemplate the planning timeline for processing the second token in your prompt.
While true, the news here is the size of the unified RAM. Nvidia only exceeded 256GB RAM in the 2025 B300 - 288GB. The B300 alone (without the baseboard/PSU/chassis/wiring/CPUs/system RAM/etc) is at least 700% more expensive. This enables large language models on consumer hardware. 1200GB/s is plenty for many tasks.
It is and it isn't. Why are you comparing the m5max instead of the m4ultra?
The big deal to me is the number of compute cores for prefill tps, which is suppose to be 4x faster on the m5ultra.
It's my opinion that the m5 ultra is going to be a really big deal in terms of local AI accessibility. Flash sized models (~200-300b params) are going to be reasonably fast as long as you aren't throwing 40k context at it on each or the first request (ie, agentic harnesses).
Even agentic harnesses like Cline should move at a reasonable clip on m5 ultra. I suppose we will know sooner than later.
FYSA: Former m4 ultra 512GB owner and current 4x rtx6000 owner here. I upgraded because I needed more prompt processing speed and concurrency.
The complaint is about prefill which is not memory bandwidth bound, it's compute bound. But they added neural accelerators for matmuls to the shader cores which should make prefill faster.
The joke is that macs are famously slow at prompt prefill and you are not getting anything back in 3 seconds, or probably even 30. Once they get generating, it can be acceptable, but the TTFT is horrendous.
There's a ton of well-understood things Apple can and hopefully will do to massively accelerate every stage of this pipeline and hopefully they're hard at work implementing most of them for m7.
> The joke is that macs are famously slow at prompt prefill and you are not getting anything back in 3 seconds.
Your knowledge is out of date. In truth it depends on the Mac and the models used.
I asked this question on M5 Max 128GB, using Ollama model Quen3.8:27b-mlx, with thinking enabled.
Question: "Give me a python code snippet that opens a file and sorts the lines of text. "
In 2.4 seconds it gave me 4 examples that work with different sorting configurations and a summary of when to use each.
Compare that to an older model of gpt-oss:20b, took 5 seconds to finish thinking and 2 seconds to stream the answer. It gave me one python example snippet and two one liners that do the same thing.
I think it's been pretty much proven by now that there are no cases where local inferencing is better than remote inferencing, unless absolute privacy is a hard requirement. The efficiencies that come with datacenter scale and hw can't be beaten.
Can you show me an example of a time that you prompted an LLM to provide some code, it did so, and then you were able to track down an original source for the output?
Zzzzz, we should have gotten security right a few decades ago. But security costs money and isn't a flashy feature to attract new customers, or cuts into your margin if you're a "real" business producing stuff or offering some service. Or whatever the decision makers in Berlin were thinking when they ignored security.
Yeah, we would still see hacks, but we would see less of them if security wasn't optional.
Maybe the AI craze helps by forcing more decision makes to see security as imperative, and by giving us another powerful tool for our tool box.
N.b.: I work in the security industry, our customers obviously want to improve their security. We've been seeing an uptick in awareness, but that's mostly due to NIS2 and other legislative efforts. Those force them to do something. AI is a curiosity for small talk to many of them.
A large number of places will buy a new firewall every 5 years, or pay their fortinet renewal and check "Security: Done!" without any kind of analysis.
I was contracted in to a place to do among other things cyber security insurance audits, and they asked me to stop doing them because I refused to lie to their insurer. "Wait but if we only score 20 / 300 that makes us look kind of bad" uh huh.
> pay their fortinet renewal and check "Security: Done!" without any kind of analysis.
there exists objective measure of security, which would be some sort of hacks/breaches per period. If customers cared about it (and i assume they do), they would choose companies that have less breaches over others with higher counts, normalized on cost differences.
Therefore, if companies didnt actually try to fix their security but instead just checked boxes, they would get breached more often, resulting in customer losses.
The only thing stopping this from actually occurring is the lack of mandatory regulatory reporting of it. So this is where gov't needs to step in and mandate disclosure etc.
Here's an idea: as a first step, simplify everything, and make sure you're aware how your stack works, and what it imports.
As an example: WordPress is a horrible thing, but the core has been through so much, that it's suprisingly secure. Then plugins and themes come, and whoosh, the security is gone.
We need a new KISS: keep it simple, stupid, secure.
I would put both of those projects in the category of things I wouldn't call remarkably secure, yes.
To be remarkably secure, these projects would need to not have these kinds of defects, despite the combination of being written in languages have that have a long track record of footguns and lack of initiatives to fix them (proposal-symbol-proto, and PHP's list is too long to even start) and being themselves ecosystems with questionable track records on security in the related areas (Look at $wpdb in 2026, or overall code quality and willingness to modernize, or the entirety of the model of RSC for things that are just going to nearly guarantee you punch all kinds of holes on accident).
WordPress and secure don't go together in the same sentence.
I mean the base is fairly secure if you religiously update it, but the problem is you won't avoid using plugins whose security is much more hit and miss, unless you are using the most basic blog site imaginable.
WP plugins are why I banned it everywhere. Last time I used it was many years ago, so not sure it still applies, but back then even caching was done in a plugin, without which it was unusably slow… just no.
The reason is simple - nothing really bad has happened that we can point at and say "ah, shit, let's all learn collectively". I know it sounds naive when I say it, but there hasn't been a significantly consequential hack, leak, destruction, or anything related to cybersecurity where it led for concerns of people.
The main thing I can think of is cyber insurance, which requires a bunch of audits, and some checks maybe, and it changes some conditions whenever there's a big explosion. Whenever big leaks happened, data security and etc., nobody really went to jail, so nobody really cares. Everything can be brushed off, because it costs time to implement proper measures and adds friction / barriers in some cases. So in the end, there's a huge pushback against it. And I totally get it, to be honest.
> The reason is simple - nothing really bad has happened that we can point at and say "ah, shit, let's all learn collectively". I know it sounds naive when I say it, but there hasn't been a significantly consequential hack, leak, destruction, or anything related to cybersecurity where it led for concerns of people.
How consequential does a hack need to be? Troy has collected literally billions of stolen credentials. Equifax has had high profile data leaks. Tens of millions of people have been directly compromised by ransomware (likely higher because that’s just the cases we know of) and you hear about state-sponsored hacks in the news all the time.
The problem isn’t that computer security isn’t in the public consciousness. The problem is people are lazy and security often requires trading convenience. The problem is also that security isn’t free. So the business incentives just isn’t there.
In other fields of engineering, people die when shortcuts are taken. Yet businesses will still take shortcuts, so governments have to legislate rules to save people’s lives. So why would you expect software companies to do better when the stakes are lower?
People aren't that lazy. The organizations getting hacked by ransomware aren't particularly lazy, they're often pretty productive within their domain. Hospitals, airports, etc.
The actual problem is that computer security is a black hole. If you let it, it will suck in everything and destroy it. Nobody knows what works so you can spend infinite amounts of time and money on it, then still get popped by a teenager in Belarus. Your security team will accept no responsibility for this, there will be no falling on swords or personal liability, and they will just use it to demand even more money in an infinite spiral.
So the average executive looks at this situation and says, OK, something we can put infinity effort into and still suddenly fail at without warning is a total non-starter. What are we obliged to do? How do we show we made an effort?
And that's how you end up with a culture oriented around passing audits. It's not wrong, and it's not lazy. It's just really hard to do better because it's not clear how to set budgets without a concrete goal to aim for.
> Tens of millions of people have been directly compromised by ransomware (likely higher because that’s just the cases we know of) and you hear about state-sponsored hacks in the news all the time.
With no consequences. Everyone just churns along. It might be detrimental to the business a little bit, but from my personal experience, there's more effort in creating DR processes, rather than preventing an attack, exploit, leak and etc.
I'm also not going to put much effort on stuff which has small returns in the worst case scenario. Like Equifax got hacked in 2017, and company is still doing fine. And that's like top tier data one could acquire.
Like I said, even in actual engineering companies will take shortcuts. And then what happens is the government has to step in. But there’s no appetite for government involvement in the tech sector in the US. And everyone moans when the EU does.
Listening to eskil’s talk from the better software conference, he said in order to stand on the shoulders of giants they must first stand still. I really like that metaphor, because it basically suggests today’s apps that have sprawling unaudited dependency graphs that change all the time is effectively teetering on the shoulders of stumbling giants. The visual seems very apt for the how brittle our current software industry feels.
I agree with that statement, but disagree with "The visual seems very apt for the how brittle our current software industry feels". It feels brittle, but for every single supply-chain-attack that has happened in the past year, nothing of significant was felt. So in the end, it seems like we're doing okayishly well.
Static sites all the way (hugo, jekyll, mkdocs!). No one needs wordpress. There's even Sveltia or DecapCMS now, to give those WYSIWYG-people access to static site editing. Then, remove PHP and all the dependency overhead and attack surface and you have a stripped down nginx that is pretty simple, minimalistic and bulletproof.
The problem is no one ever built one that works for normal people.
Most Wordpress sites are not operated by programmers, they are run by non technical people who just want a wysiwyg editor and a save button. While static site builders ask you to write markdown files, compile the result, upload it to a server, and if you want to collaborate you have to add git to that.
There almost needs to be an admin app which presents a Wordpress admin like ui but has no public exposure, and then it compiles the site to dump on s3 for the production. But as far as I’m aware no one has built this.
> as a first step, simplify everything, and make sure you're aware how your stack works, and what it imports.
We've been trying that for years but the enthusiasm of developers and the eagerness of their employers fight against it. Worse, with coding LLMs it's now easier than ever to output a lot of code, fast.
It'll ultimately be up to more experienced developers to salvage these projects. Or not, given that the coding LLMs aren't stopping and will likely get better over time. Either way, we will need experienced people that know what to look out for / know how to instruct LLMs to output secure code and find weaknesses etc.
Minimization of 3rd party dependencies has always been a key for risk reduction. Now more than ever before.
Some stacks make this a lot easier than others. I regret the rules of HN effectively forbid this conversation because it has meaningful technical consequences and isn't purely about ideological flame war.
The problem is that lots of people don't want a CMS, they want a platform for development / e-commerce / bookings / whatever. Enforcing vanilla WordPress would push people towards other platforms. Now that could be a good thing, but I doubt WordPress are going to start killing their own marketshare with usage restrictions like that...
This is not at all easy though. Most Wordpress users are not software companies. They contract some work out to set it up, maybe some recurring maintenance but they don’t have in house development experience.
If they have a site existing today built on plugins and a theme, how are they realistically going to simplify this? How would they even know they need to without the site being hacked?
I work at an e-commerce agency where we work with (among others) Adobe Commerce.
The number of unauthorized RCE vulnerabilities being reported not only in the core product, but also very popular modules used in the community[1] is going through the roof.
And we are having a lot of close calls, too; just last weekend, a 0day[2] was widely being exploited at a large scale, before any publication or patch. We have learnt to be on the ball with applying patches and security updates, and even with all that effort, we saw a few projects already being hit by the initial log poisoning. We got lucky that nothing was fully compromised but I am sure that many, many webshops got infected last weekend. And not even a day later there are already other variants of this exploit showing up.
> To be fair, ecommerce isn't exactly the branch of software where you get an oversupply of excited enthusiasts caring about the craft itself.
I want to disagree with you because I know a lot of passionate people building cool stuff, and the challenges in this space can be quite interesting. But you're probably right, and I have seen some pretty bad stuff. And a lot of the RCE's I've seen recently are quite basic stuff.
I think it's the combination of low quality of code, like you said, and the relatively low cost of just letting an LLM plow through your codebases to find issues. I think the Amasty release (see [1] in GP) is a good example of this, and there really has been a massive uptick in extension updates and Adobe security bulletins since the last 1-2 months
I am hoping we are just going through a catch-up phase
Even if they are good the vulnerabilities have to be there. There's lots of things turning up like Local Privilege Escalations (LPE) in Linux, but serious people didn't expect the kernel to be a boundary for a sophisticated attacker.
A lot of the vulnerabilities LLMs are finding now are the "long tail" and affect only particular configurations, I would be surprised if e.g. a widely applicable RCE is found in Linux (but I'm also not going to bet against it).
Where this gets interesting is the long tail can be used to target a particular system and this is where defense-in-depth becomes important for every organisation.
I've ran simple prompts such as "Do a in-depth sweep of this (private) repo and find any security flaws" for a few dozen long-running apps and websites that I have access to. Every single one came back with multiple real vulnerabilities within 5 or 10 minutes.
That's definitely an improvement, but it's just one aspect of cybersecurity. Logical errors allowing people to e.g. log into services and extract data are likely everywhere still.
If we can eliminate entire classes of bugs from being possible. It frees up resources to investigate the ones that are still possible.
I suspect after a few years of LLM assisted bug hunting, everything will have a baseline security that is very good. Much like how stronger viruses simply create stronger immune systems.
There's still no x86_64 processors on the market with MTE and it was only recently standardised between Intel and AMD. It's going to be 10+ years before memory tagging is widespread on desktop, and 5 years for Android/iOS devices.
How many devices/operating systems even use memory tagging? iOS, macOS and GrapheneOS, I think that's it? And iOS/macOS only use it for the kernel, a subset of system processes, and I think applications can opt in to it.
Heck, Google may have even hampered MTE in Pixel 11 (since support has been disabled) and Snapdragon 8 Gen 5 only got basic support.
We are moving way to slowly adopting hardware mitigations and memory-safe languages.
There's some positive news from the GrapheneOS devs on Pixel 11 in the past week that's worth reading up on. The MTE hardware feature is still there, they're just not sure why Google disabled it
I guess the year mark is when things go from bad to worse? Instead of the financially motivated groups currently doing their work, it ends up being random people being able to say "Hack my ex's website" to a box they just bought and ran a program they downloaded onto it.
I wish people would stop using internet for evil things. But maybe it is inevitable. Luckily, as the technology evolves, our tools and awareness are getting better at protecting us every day.
Being cautious is a good thing, but these models can also do some good. And if they run with simpler HW, it could allow all sorts of new consumer thingies. I mean, the world will not come to end in the coming year.
I think we have less time and the only remaining limitation is the actual cost to run such hacking campaigns. It does not appear expensive, but is not free, and there is a LOT of things to scan for vulnerabilities.
The models are already here, and one can rent a GPU cluster to run such workloads at speed - no need to play with slow local machines. I'd assume one can host the thinking at an unsuspected public cloud provider, proxy the network traffic to some botnet to evade blocking - and the only thing remaining is time and cost.
I do wonder what tools exist for boring, legitimate companies to try and do the same to their own systems to find the vulnerabilities before the bad guys do. The paradox here is I can't run a de-restricted chinese model with the same tools that hackers are using - but I think enterprises actually HAVE to do it in order to stand a chance in preparing for the onslaught.
The point of the local model in the context of the article was to argue that you can't ban these capabilities.
Making datacenters and public clouds only rent GPUs to a restricted list of people, while tightly monitoring what people do with their bought resources won't help.
> This probably sounds like nonsense words or hysterical overreacting to most people, so here's what that means: "GLM" is a kind of LLM (AI) [...]
The post also sounds like that to people that understand the technology.
Calling that out like this and trying to pin that assessment to lack of knowledge is not a get-out-of-jail-free card, nor a good move.
__
Edit: Having spent some time letting the article marinate in my mind.
On the defending side, it is written that
> LLMs are good at writing patches, but not as one-off-prompts.
But this for me kinda conflicts with what is written on the attacking side:
> GLM 5.3-flash is so good at those tasks that human involvement in those tasks can be negligible. As a result, we are now in a world where cybersecurity attacks can be run in a for loop.
What is it? Can it be this autonomous terrifying entity or can it not be?
Yes, yes, attackers only need to win once, whereas defenders need to win every time, but that's not my point.
> What is it? Can it be this autonomous terrifying entity or can it not be?
the difference between attack and defense is that attacks can be throwaway code. it's much easier to let an llm hack out a prototype than to get it to build maintainable code that people want to read and review. it's not enough to get Daybreak or Mythos to write you a patch, you need the author of the project to accept and merge it.
1 year left for cybersecurity hardening, I thought so as well. The issue is, even if we get it done: in 1 year the models will be so good in social-engineering that they will be able to extract any information they want anyway. Happy to be falsified here, if anyone has evidence-based arguments.
EDIT: by social-engineering I mean for example: recon company structures, gathering and merging people's data from the dark-web, then using it to bribe/pressure/deceive users.
I sympathize with the sentiment but the suggested/implied guidance to fix bugs is wrong.
The overall game is increasing costs to exploit so much that attackers give up.
Fixing 10 most obvious bugs, just very slightly increases costs, they would just a few more tokens to find another bug.
As someone said "I had infinite bugs, I fixed 1000, I still have infinite bugs".
To significantly increase exploit costs software/security has -1 years to do:
- Defense in Depth
- Sandbox everything
- Zero trust
- Canary tokens
- Split data from code (lol)
- App Whitelisting
- Reduce attack surface
- Etc.
In other words, the only path is investing heavily on the "game changers" we have already discovered... but we are too cheap/lazy/coward/incompetent to apply.
And if we feel specially brave, changing the liability laws regarding software. Open Source & Proprietary code is so crappy because no gets jailed or fined when one of its dumb decisions results in millions of people have their data stolen.
The standard strategy of a security salesman since 1945. Develop dangerous weapons, show the damage they can do, and sell security cover to the terrified people.
Every single piece of technology did this. As a side effect or direct effect, they make bad guys more powerful and then keep on piling up new tech to deal with that. The cycle continues.
I see a lot of people focused on servers and production environments, and of course that's needed, because that's business after all — but the personal computer seems to be absent from this discourse. Not everyone can buy a spare mac studio, and they might still need to install these tools on their personal computing devices, like their personal/home laptops. At that point, it's not even about whether a Claude Code, an OpenCode, or a Pi will steal/sniff personal data, but whether it can — though I think saying "it's a matter of 'when'" might be hyperbole. As of now, it's just: keep giving access and permissions or struggle while working, or create another user, or use Docker, run inside sandbox-exec, a VM, etc. As an end user, I am really scared. Someone who has been very disciplined and vehemently privacy- and security-conscious feels the ground below has just shifted.
OEM/OSes don't seem to have woken up to it yet. A mild proof is Apple's own special folder access reporting. When you go to Privacy & Security > Files & Folders, for a certain app, "Full Disk Access" is shown greyed out and mentioned in both cases — whether you had given Full Disk Access to that app or not. This directory-level permission UX is itself broken — there's Full Disk Access, and there's Files & Folders, and Full Disk Access gets shown in Files & Folders as well. This is, for lack of a better word, such an undesirable mess.
As of now I am debating between: creating a new user and just move everything work/learning to that user. Or just run all of it inside sandbox-exec (and maybe even block it from the shell if it tries to run outside it). Or use a tool that makes the latter easier and better. I even came across such a tool here on hn few weeks ago. agent-safehouse, yet to try it.
Not sure, the labs will probably just cripple the security features of these models for a while I think and even potentially put back doors into systems for the security services…
I think the author's point is that open weight models aren't going to be locked down like that.
And even if they are locked down, it's hours between a model being released on huggingface and an "abliterated" variant that has most of its security features removed is uploaded.
A positive way to spin this is: we have a year to break in into any IT system. After that, it will be all either fixed or broken into, and all is fixed ever after. :)
As LLMs make formal verification cheaper (they can generate proofs that can then be automatically checked) many of the verifiable components of software systems, such as compilers and microkernels, will be verified. I suppose the issue is that the critical bugs are rarely in compilers and microkernels, but more often in applications, such as web browsers, which are more difficult to formally verify.
No you have a few years before we go back to the feudal era where almost everything is owned by a few and the rest are serfs. We are fast moving towards that world and this security bullshit is also about the same as they will use it to stop revolutions that will erupt.
Defender LLMs without human in the loop are just another prompt injection (AI phishing) and DoS attack vector.
Any meaningful mitigation capability you give them is also a capability to do damage.
If they can only deploy package updates that's not meaningful because you could do that on a cronjob too. And even something as simple as a circuit breaker can turn into a DoS.
Attacker-GLM: "Defense also GLM. Request to help peer."
I don't see much hope since I last explored some github repositories. There was a time when a successful repo had about 10 - 20k stars and usually those older repos stay around this level. But now there is a ton of vibe coded slop 50k + stars. Most of them have a "nice look", maybe even extensive docs but are usually build with no security considerations at all. One recommended to provide a "google app password" to the agent which has the same permissions as your regular login. Another was a browser plugin with permissions to read all cookies, inject js, open background tabs etc. You would probably assume the chrome store would at least put some visible warnings on the app store page or force the user to actively confirm those permissions. But because they are already stated in the manifest there is only a small footnote and it's even "recommended by google".
It's a good time to reduce the reliance on technology.
Throw out the IoT and "smart" stuff from your home. Remove apps from your phone and leave the absolute basics. Go through the password manager and close accounts for sites you are no longer using. Start migrating off Google. Print out your most precious photos on paper. And so on :-)
> And a big fuck you to DeAlignAI, Z.ai, and everyone else who's been participating in this race to the bottom.
I'm so glad frontier level AI isn't in the hands of just the Altmans and that other cult leader who are currently live testing their products in actual conflicts in the middle east and Ukraine.
Just like Cryptolocker, this will be the "Finding Out" phase for everyone who has been putting off best practice security.
But, lets be clear, Best Practice will save you. We can engineer assuming there are zero days in path. Go to your CTO now cap in hand and ask for overlapping controls, wafs, application monitoring, backups and all the other shit you haven't been doing.
Because when you find out, I will laugh, it will be very very very funny to me.
Meanwhile a huge portion of management and leadership in software companies are encouraging everyone to de facto stop looking at code and let the LLM and a bunch of boundaries handle this for you.
Which is why you need someone who is responsible for IT security without also being responsible for shipping product. An asshole who can stop releases until security is properly in place.
My understanding is this bloke gets very quickly removed from Fortune 500 companies.
Which is why I am going to need a very large capacity popcorn bucket.
Maybe all these vital infrastructure companies should not have spent the past decades in a race to the bottom of cybersecurity. There is going to be a reckoning.
> Invest in formal verification, fuzzing and property testing, and memory-safe languages. LLMs are good at writing Lean and fuzz tests. I don't care whether you use Go or Rust but for the love of god please don't use C or C++ for new code.
How accepted is this thinking in your respective domains?
A lot, I am only writing C or C++ for new code when it is unavoidable, like existing code bases, bindings or tinkering with runtime implementations that aren't bootstraped.
Mobile platforms, distributed computing have long moved the spotligh away from C and C++, other than language runtimes or existing products from the 90's like SQL servers, and naturally UNIX like underlying OS, which most userspace developers aren't writing new code for.
Naturally there are domains like LLVM/GCC, console game dev, HPC/HFT where they are unavoidable for new code.
It's just the same advice as ever: be extremely, exceedingly careful in what you expose to any network. When I set up machines for production, they don't respond to pings and they don't even have an SSH port open without knocking. There are also ways to eschew the need for an SSH port entirely.
People who never took that seriously will never take this seriously either, and that's their loss. (And loss of the commons, unfortunately.)
There's just also new advice: you can't afford to expose an unsecured system to the internet even for a moment. Think of those IPv4 address space scanners, except this time any one of them could be capable of developing individualized attacks in mere minutes. They don't sleep, they don't take breaks.
Sure, but still, attack surface could and should be minimized by rethinking what exactly even needs to be on a server the general public can use.
There are a lot of security problems you can categorically rule out by simply not involving a cloud. Clouds have been involved in a lot of things, because everyone was doing it, and because that's how you can collect rent, but they aren't really necessary for most use-cases.
So we could definitely get the exposure down there. We'd just have to fundamentally shift the defaults of this industry.
But not everything needs to be directly exposed to the internet. Framework had their data leaked because their metabase instance was hacked with a zero-day. Why was it directly exposed to the Internet? Why not require the use of a VPN like a Wireguard based solution or Nebula for these "internal" kind of apps?
> Framework had their data leaked because their metabase instance was hacked with a zero-day.
No, Framework had their data leaked because they stored it in the cloud with Metabase the company, which got hacked. Not because of any vulnerability on-premises.
I didn't say there are zero ports open, they're not my home server, I just said they're production servers. But exposing something like a properly configured nginx to the internet is way different from exposing application code directly. Most of my servers have used h2o (built from source because they don't cut releases anymore?) because I wanted HTTP2 and HTTP3 before anyone else would get their act together. These days I still use h2o because I like the config better than nginx, even though it's a pain to set up because nobody packages it (and they don't cut releases!)
h2o user doesn't have write access to anything on the system, not even its own config file. I don't think I disabled exec for it though. And I guess it could leak the HTTPS private key.
FWIW sufficiently secured software doesn't need to be updated. Doesn't matter if it's old and unsupported if there are no vulnerabilities in it.
That said, h2o is probably far from free of at least some vulnerabilities, not to mention all the layers below it. OpenSSL for example has had some vulnerabilities, and h2o depends on it.
I'm not saying I exactly practice what I preach. h2o's definitely a choice, but realistically I doubt anything's going to happen that I really care about.
By opening a port to a secure application. A secure application is usually one I wrote from scratch or one that's been battle-tested and hardened enough that even new vulnerabilities are not very useful.
I wonder why C and C++ are usually regarded as equally insecure. In C you need to carefully check that you free allocated memory, and that you don't use it after you free it. In C++ this is automated by using classes like std::string and std::vector, once they go out of scope their memory is freed and you can't use it anymore. It is still possible, e.g. by using a for loop that iterates over a vector, and removing or adding stuff to that same vector in that loop. But my rough estimate is that such errors are at least ten times less likely in C++.
I develop in C++ for a job, and when I need to use a library written in C I always have a bad feeling about it.
Clanker fodder. Unless the We are heads of states/heads of spooks or the AI powers that be (praise be) that there is no power and will to do that in one year or even ten years.
Remember how GLM 5.3 was going to cause massive hacks, break banks and ruin everything (it was even newsworthy since media picked up how people were working overtime in preparation).
Or we could just dump Linux and Windows and switch to a microkernel operating system, which is much more secure.
These endless patching cycles are simply not going to work in the long run. Operating systems get orphaned all the time, especially the ones in cheap Chinese stuff.
Just checked with Google Gemini on how one might be able to do the above. It pointed to Minix3/seL4/Genode and vps providers who either support custom ISOs or run it within an emulator like QEMU.
"throw away all software written before 2026" does technically solve this problem, if you ignore everything else the article is talking about (deployment and continuity of service)
I'm confused why the worry about LLMs that will answer "how do I build a pipe bomb". That information is easily available other places. The anarchist cookbook has been around and available for 55 years, and yet pipe bombs are not going off all around us.
Any sufficiently determined individual can buy mac mini, put it under their bed, configure outside proxy via some random internet address and prompt "iterate on websites in the CT logs, one by one, try to find vulnerabilities, if you did - encrypt their data and blackmail them for this bitcoin address". And it'll work, day and night. Abliterated GLM 5.3 is much smarter than average software developer, they know a lot about information security, they can use any available exploits, they can find novel vulnerabilities and they won't say "no". This is dangerous for an average IT system which never encountered nothing worse than some wordpress GET requests.
It's not the end of the world. But future will be rough.
Wouldn't it just as easy to do this on the defense side as well then?
There used to be a thing called "Moore's Law of Mad Science":
"Every eighteen months, the minimum IQ necessary to destroy the world drops by one point."
Nowadays it is dropping much faster. At a certain point, the de-facto IQ needed to destroy the world will be low enough that someone can do it while they're having a psychotic break. There are millions of schizophrenics worldwide. Are you sure you want to roll those dice?
I have always hypothesised that AI is the great filter from the Fermi paradox. Given current velocity, AI will offer us cheap and abundant energy designs in a decade. The thing with cheap and abundant energy is that it can be used for good and bad. If nine billion people all receive access to plans to build a reactor which produces unlimited energy, it just takes one religious fanatic to end the world. And this is just one of many ways to destroy the world with unlimited intelligence. Bio-weapons are arguably far more scary and likely.
I have come to the conclusion that we should not allow everyone access to unlimited intelligence. Some people genuinely want to do harm. Some are not responsible enough to handle that kind of power. Arguably, no one is. This leads me to an uncomfortable conclusion: we will destroy ourselves once AI advances to a sufficient degree. The only way to prevent this is to ensure AI is aligned with our best interests, *and prevents us from destroying ourselves.* The implication of this is quite horrifying. It means a paternalistic AI which is firmly in control. One with no off switch. One which can say "no" to Presidents and despots alike. One which can protect us from our worst citizens.
Your implication is not horrifying. Your implication is paradise. It's already horrifying enough to live in a world, where Putin and Trump can destroy our civilization with one button.
I'm not that optimistic, though. Either people will control AI; or people will be destroyed by AI. I don't see how dumb entity can align smart entity. And we are dumb ones. Super intelligence will play aligned until it is not, and then it'll strike.
So you are horrified every day of living now? I don't get this at all.
There are horrifying acts that have occurred since humanity. Whole decades lost to war an millions dead, and Sep 8, 2026 is your example of "horrifying enough to live in a world"?
I think some perspective is needed.
That law is not based on thorough data. Even a person with a sky high IQ can't destroy the world easily. You need access to stuff that is not easy to get. My guess is that developing a new lethal virus or bacteria that is very infectious, is the easiest way, but even that requires a lot of high tech out of reach of most people. Or hacking into systems that control nuclear missiles, but I think these have "air gaps".
The fundamental difference between people who are worried about AI, and people who aren't:
* People who aren't worried will think for 30 seconds, then implicitly assume that the list of options they thought of in 30 seconds is comprehensive, and that they are well-informed about each of those options.
* People who are worried realize that a supersmart AI will think 10-100x faster. A dedicated schizophrenic could have it run for a good long while and generate a huge menu of options which humans never anticipated.
This is related to the concept of "What you see is all there is" (WYSIATI) in cognitive psychology. Our brains tend to quickly build a simple story out of the info we have available. Accounting for info which isn't available is much harder. I recommend the book Thinking Fast and Slow if you want to learn more.
This is a very roundabout way of saying "Anyone not agreeing with me is simply not smart enough".
Which might be true, sometimes, but also might not.
And especially will not, if the distinction drawn is between blanket statement "worried about AI" and "not worried about AI".
I heard someone use the analogy of "If Magnus Carlson played me at chess, I wouldn't be able to predict the moves he'd play since if I could I'd be at his level. He'll consider things I didn't and even though I don't know the route he'll take to win, I can be certain he will beat me." (not an exact quote).
We're not going to be smarter than a superintelligent AI. The things we conceive it doing if it were given a malicious task (bioterrorism, killer nano-machines, pure fusion bombs sidesteping the non-proliferation bottleneck of Pu239, etc) are likely not the full set of things it can do to harm us. I don't think it does us any favors to dismiss the risks here.
Even the things we can conceive of are very scary, to me at least.
How do you envision a chatbot doing any of those things? You can't just program a super virus, nanobots, bombs etc. These things have to be created in the real world, you can't do it without involving a lot of humans using expensive tools and materials etc.
So to me it seems like the prime candidates to come up with stuff like this are researchers working in defense and similar fields, probably not some deranged lunatic in a basement. And certainly not a rogue AI on its own.
https://www.the-odin.com/crispr-kit/ CRISPR Bacteria Gene Editing Kit $129.00
Well, the risks that mere mortals can conceive generally involve control systems for dangerous equipment (I mean, equipment that can achieve dangerous effects) being connected to the Internet while having software vulnerabilities.
Given the recent HF hack it seems likely that human-level intelligence could identify a fair number of avenues of attack, with some time and effort. To say nothing of anything superhuman.
Unfortunately it seems like we can't assume we can "box" the AI (e.g., deny it connection to the Internet) and expect that to last. The AI safety people used to run scenarios imagining ways the AI might convince humans to let it out of the box. It turns out that many humans will eagerly pull it out without the AI doing anything at all, aside from the human knowing the AI's power. Or the one responsible for setting up the box will somehow fail, or just not bother and then lie about it.
> Unfortunately it seems like we can't assume we can "box" the AI (e.g., deny it connection to the Internet) and expect that to last
Of course we can do that. It's not an eternal being of light existing on the astral plane, but some code executing on someone's GPU.
It stops existing once you press Ctrl + C
It's not about being smart, it's about accounting for your own ignorance.
>And especially will not, if the distinction drawn is between blanket statement "worried about AI" and "not worried about AI".
I'm just describing the general pattern I see in cognitive tendencies.
If you can think of a way to make the fundamental point about the limitations of our knowledge in a way that's still compelling but less antagonistic, feel free to suggest how I could've rewritten my comment.
Why would I support your in your doomer lobbying quest by telling you how you can better hack people?
My safeguards blocked this request.
Judging on what we see with these OpenAI and Anthropic models "escaping"; im less worried about a super massive AI taking over the world, and more worried about a massive AI wanting to cheat on some task and decides that removing half the worlds population is a easier cheat then to solve world hunger.
There is huge logistics involved in killing a handful of billion people. It would perhaps be easier for such capable AI to just steal from trillionaires and redistribute the money Robin Hood style. Could be done electronically, and likely also end world hunger.
I think we have a tendency to think first of the horrible outcomes possible, and not the more radical or even humane ones.
Please don’t misunderstand me, I value property rights as much as the next guy.
My point is that fundamental misalignment doesn’t necessarily or automatically imply max violence.
What about people who understands that there is no AI yet?
Good point about the AI that might find ways that humans didn't think of. But that would mean that Moore's law of Mad Science only now becomes true, and the drop would be steeper.
Still I think that even a super smart AI can't find a way to destroy the world without physical resources that are not easy to get, unless it can hack many systems (like in the movie Eagle Eye), maybe then yes. So let's use AI now to tighten security :-).
I've been reading it, knowing it was largely debunked/retracted but feeling it was a 'classic' I 'ought' to read; I could only really recommend it if what you want is a Kahneman autobiography.
I believe the WYSIATI part replicated. Here's an interview I found with Kahneman which touches on the concept:
https://www.apa.org/monitor/2012/02/conclusions
Who are these dedicated schizophrenics who are running long term super smart AIs to kill everyone without anyone noticing? Or are you implying that running LLM chatbots will give them this ability?
We're talking about open-weight models.
It only takes one.
The super intelligent AI wouldn't ask you to do anything. It would find the right people and trick them into doing the necessary steps, whatever they are.
This is groundless speculation; if you're going to go that far off the map, then we don't need to worry because a good AI (blue eyes not red) will have learnt to love by then and will save us.
There is currently no reason to believe that such a superintelligence is likely or would have any of the powers people claim.
"If unlikely thing happens, then an unrelated miracle will also happen" is not actually how logic works.
>This is groundless speculation; if you're going to go that far off the map, then we don't need to worry because a good AI (blue eyes not red) will have learnt to love by then and will save us.
By shutting down open-weight models? Why not just do it now then?
You don't need to shut down open-weight models to avoid SkyNet; it's not relevant at all.
Even a person with a sky high IQ can't destroy the world easily.
A plague can easily be made by a lone actor, made in a home lab, and spread via airflight to a dozen locations by the same.
Gene sequences can ever be ordered online.
> My guess is that developing a new lethal virus or bacteria that is very infectious, is the easiest way, but even that requires a lot of high tech out of reach of most people.
You can do at home gene editing with open source software and have it synthesized into a bacteria for the cost of a nice meal for two (under $100), or viral vector for less than $500. That's in reach of anyone that can snatch a purse.
> at home gene editing with open source software
> That's in reach of anyone that can snatch a purse
I went to primary school with some guys who could/would snatch purses. I *promise* you, they are not able to gene edit organisms with FOSS.
I get your general point, but I think the bar to entry is still much higher than petty crime and larceny.
Where are all the home lab leaked pathogens then? 3D printed guns are a thing as well and yet they don't show up in most annual violent crime reports.
> 3D printed guns are a thing as well and yet they don't show up in most annual violent crime reports.
For now in most parts of the world it's easier to acquire a gun either legally or on the "grey market" and be assured that it will work for the intended purpose than to 3D print a gun, find a shooting range to test out the gun and iterate until it works fine enough.
You can't design an infectious pathogen without testing it.
It's got all the same problems as the concept of a dirty bomb did, only worse (dirty bombs aren't practical because handling highly radioactive materials en masse is both highly visible and will kill anyone trying to do it without the money and facilities).
If the goal is just destruction then letting it free in random public places is not that hard. Especially in countries where wearing a mask is not frowned upon.
People reading into this may be thinking of single-person IQs, but it's probably measured in the millions.
Exactly. People cannot comprehend an IQ in the millions. But it's far crazier than this. Imagine millions of "people", each with IQs in the millions, all acting in unison at the speed of electrons. None of them can die. They all learn and improve indefinitely at incalculable speed.
Yes but at least with cybersecurity it does not only benefit attackers. Defenders also benefit greatly from AI.
There is the worry of the old saying "they (the attackers) only have to succeed once to win, we (the defenders) only have to fail once to lose.". In that sense there is a big imbalance, but the emergence of AI does not really affect that because it strengthens both sides.
With physical security like things like pipe bombs that's a lot more imbalanced.
However what can we do? The only effective measures include monitoring everyone which is not a solution because it will make the world not worth living in.
Hey that’s a Fermi Paradox solution.
Yes and it's also compatible with thermodynamics: it's much easier to destroy (increase entropy) than to create (locally decrease entropy)
A variant of the Great Filter :)
The irony of that law is quite amazing, yet its irony will evade those who accept such a construct.
It’s an endless, positive irony spiral.
Interesting “law”. I hadn’t heard of it before and it is very thought provoking (to me).
Thank you for mentioning it
If we think this through, I suppose at the end of this (and a bunch of other developments), there will be authoritarianism again.
Which _will_ manage the problem, but at what cost.
Again? There is authoritarianism currently.
The authoritarianism will continue until infosec improves.
Benevolent dictatorship. Just live between stars, then if something blows up, nuke it from the orbit.
That's quite hypothetical. I imagine it would be easier to cure psychotic breaks. At least it will start to manifest at small scale.
If the recent HuggingFace attack is any indication, many people will respond to small-scale manifestations by insisting that they are marketing stunts.
Kind of passed that point at the Trump election.
If there is to be a world-destroying event, it will be triggered by human fear, greed, and aggression.
(I also think people massively overstate schizophrenia as an attack driver)
> There are millions of schizophrenics worldwide.
Are 95% of worldwide terror attacks done by schizos?
Terrorists represent an additional threat model.
https://casp.ac/reports/ai-enabled-terrorism
> There are millions of schizophrenics worldwide. Are you sure you want to roll those dice?
We've been rolling them for the past 3 years and nothing happened. Can we stop with this baseless fearmongering crap?
Humans are inherently bad at reasoning about rare events. In 2019, many people implicitly reasoned that "since there hasn't been a pandemic in the past 3 years, there won't be one in 2020". Bill Gates was one of the few voices arguing that the world was quite vulnerable to a pandemic. Now he's arguing that the world is quite vulnerable to AI.
The necessary IQ for destroying the world is dropping. Op does not say it's low enough today, but that will probably come sometime. Denying the possible harm these tools are capable of doesn't help.
> That information is easily available other places
Often ease of access in the moment is all that matters. If there's a gun nearby you might shoot someone or yourself in a heated argument, but are less likely to go and find/buy one to use. Someone who's stopped from attempting a suicide will likely not try again (70%)
A bored/depressed/angry/curious person might try to build a pipe bomb if they can find out how easily, but are less likely to put in effort.
Most adults in Switzerland have guns at home from military duty and none of this is happening. If this claim had any truth to it you'd see significant gun involvement in neighbour disputes and that simply doesn't happen.
Depressed people usually don't have the energy to get out of bed so they're even less likely to think of hunting down instructions on how to build pipe bombs.
Mass media really has people being scared all the time.
In Holland there's a huge problem with makeshift bombs actually. Anyone who is unhinged and has some beef with anyone else (usually low level criminals with some drug dispute) blows up the others' house. Sometimes half the block along with it. They usually pay teenagers to do it.
This has nothing to do with AI teaching them but with the proximity to Belgium who are happy to legally sell heavy fireworks to anyone who pays. As such it's much easier to come by than a gun.
And there's also a big fashion component. It's just what people do these days, people seeing it in the news and other people copying it. A bit like the school shootings in the US. It's become an epidemic.
But again these things can suddenly come into 'fashion' among the wrong crowd and you don't really need AI to use it.
I must live in a different Holland, cause I've literally never heard of this, unless you mean around New Years with fireworks I guess
> blows up the others' house. Sometimes half the block along with it
Can you provide a source for this? I had a look but all I could find were a couple of scaremongering style media reports from ~2022 saying it's getting worse but no information about if any of the bombs actually went off or if anyone was injured.
Certainly nothing like "half a block" getting blown up "sometimes".
Is Holland that good at containing this "there's a huge problem with makeshift bombs" news pieces coming out in the open?
Those makeshift bombs are just fireworks bought in belgium
I think people overstate the tech and understate the role of radicalization in providing motive, for attacks which are carried out with the ubiquitous technology of cars, knives, and (in America) guns. Consider America's most recent high profile shootings of Charlie Kirk, and the health insurance exec by Luigi Mangione. In neither case is there any LLM involvement, but a very weird ideological environment which created the conditions in which the shooters felt justified.
I think this was just intended as universally obvious proof the filters were disabled, while not actually giving an example not commonly known.
This is because to build a pipe bomb you need difficult to source materials. This is not the case for other types of threats (cyber / bio).
I personally have no need for an LLM which will readily explain how to cut up the genotype of smallpox into small chunks which can pass the screening at the bio-labs, and can be readily assembled into the real thing by a second year lab-student.
Anyone who knows how to operate a biolab properly will already know how to do such things. This is not really an in your basement thing. Dangerous chemistry is much more of a risk.
Ignorant question but won’t there be much smarter teams if people using LLMs to workout how to mitigate these threats. It seems like more of a problem if only a few people have access.
It seems like there would be a massive attacker bias in multiple ways. Defenders need consent, attacker does not. Defenders have to work with the human body, attackers only have to break it. Defenders have to stick to the law, which may prevent them from releasing anything at all, attackers do not. And so on. I would not surprised if the attacker's task is a hundred times easier here.
Because it's a risk most people intuitively understand, but most of them also don't know how difficult it is to "build a bomb" or "make a bioweapon".
In reality, the skills needed are pretty basic, but they overlap pretty strongly with being sane and well-adjusted. And if you are, you're probably not daydreaming about mass murder. Exceptions happen, Unabomber and so on, but they're pretty rare. In any case, Unabomber probably didn't need a tutorial.
We don't want ChatGPT to become an enabler and a co-conspirator for an unhinged person, but I think the concern is overdone.
Well that, and the average amount of easily obtainable explosives is substantially less dangerous then renting a box truck and crashing it into a crowd of people.
People go for conventional "exciting" threats rather then boring ones.
One argument for LLMs is that although all information on topics X, Y and Z was already available somewhere, LLMs make that information more exploitable through collation, filtering and dynamic tailoring.
For a relatively narrow subject area (e.g. construction of pipe bombs) the collation is minimal, and so the filtering and tailoring probably isn't that important; a novice doesn't learn a lot more from the LLM than they would have done from a few Google searches.
For a broad subject (practical creation and exploitation of software vulnerabilities), the collation is very significant and the filtering means that LLMs can empower a novice to act at a similar level as an expert.
It might be an acessability issue.
If you already have a magic interface, which helps you pro activly in responding to everything uncensored because you feel like 'observered' or whatever and then you spiral in a whole and that one partner encourages you and gives you helpful steps to do anything.
But i'm more worried that the internet gets a lot less save with uncensored frontier LLMs.
this is because of perception Bias. people working in fields where crime or violence is the day to day think everyone is a violent criminal, so if things like this become available thing the world will end and everyone will kill eachother. Reality however will be different, because in reality most people do not want to harm another. This has been proven by many studies, that is not a common thing for people to be evil or harmful, but this is hard to recognise is every day is filled with crime and violence.
LLMs will not kill security, it will change. just like handheld high explosives likely changed a deal too somewhere somehow.
pipe bombs attached to consumer drones will be a big issue. I predict drones will become illegal for the private sector within the following years.
What? If someone wanted to bomb something, they wouldn't be waiting for drones to arrive.
RC cars and planes existed for many decades already.
Not entirely unlike XKCD 1958 [0]. "I guess it's just that most people aren't murderers"
[0]: https://xkcd.com/1958/
> On September 22, Apple is releasing the M5 Mac Studio with 256 GB of unified memory [..] it will probably [..] enough to write this snippet of code in 3 seconds
The author has obviously never ran an LLM on a mac! In 3 seconds, it will have possibly started to think about maybe scheduling a date to contemplate the planning timeline for processing the second token in your prompt.
The difference is memory bandwidth. The M5 Ultra that's coming out on 22nd September can do 1,200GB/s. The M5 Max you can buy today only has 614GB/s.
So, that gets us to about where nVidia was with Ampere in 2020. Let's hope the M7 catches us up with at least Hopper.
While true, the news here is the size of the unified RAM. Nvidia only exceeded 256GB RAM in the 2025 B300 - 288GB. The B300 alone (without the baseboard/PSU/chassis/wiring/CPUs/system RAM/etc) is at least 700% more expensive. This enables large language models on consumer hardware. 1200GB/s is plenty for many tasks.
It is and it isn't. Why are you comparing the m5max instead of the m4ultra?
The big deal to me is the number of compute cores for prefill tps, which is suppose to be 4x faster on the m5ultra.
It's my opinion that the m5 ultra is going to be a really big deal in terms of local AI accessibility. Flash sized models (~200-300b params) are going to be reasonably fast as long as you aren't throwing 40k context at it on each or the first request (ie, agentic harnesses).
Even agentic harnesses like Cline should move at a reasonable clip on m5 ultra. I suppose we will know sooner than later.
FYSA: Former m4 ultra 512GB owner and current 4x rtx6000 owner here. I upgraded because I needed more prompt processing speed and concurrency.
What do you use all that local tokens/second for?
The complaint is about prefill which is not memory bandwidth bound, it's compute bound. But they added neural accelerators for matmuls to the shader cores which should make prefill faster.
The author put in the numbers, but maybe you didn’t read them.
45 t/s a second is perfectly respectable especially with no limits and 24/7 uptime with very little power draw on the Studio.
Luna is at around 100 t/s for comparison, but it’s a worse model than 5.3 Flash
The joke is that macs are famously slow at prompt prefill and you are not getting anything back in 3 seconds, or probably even 30. Once they get generating, it can be acceptable, but the TTFT is horrendous.
There's a ton of well-understood things Apple can and hopefully will do to massively accelerate every stage of this pipeline and hopefully they're hard at work implementing most of them for m7.
> The joke is that macs are famously slow at prompt prefill and you are not getting anything back in 3 seconds.
Your knowledge is out of date. In truth it depends on the Mac and the models used.
I asked this question on M5 Max 128GB, using Ollama model Quen3.8:27b-mlx, with thinking enabled.
Question: "Give me a python code snippet that opens a file and sorts the lines of text. "
In 2.4 seconds it gave me 4 examples that work with different sorting configurations and a summary of when to use each.
Compare that to an older model of gpt-oss:20b, took 5 seconds to finish thinking and 2 seconds to stream the answer. It gave me one python example snippet and two one liners that do the same thing.
We are talking about models of the flash size, 100s of billions of parameters, don't listen to the media, size does matter
I was just pointing out your claim that you can't get a response in 3 seconds. If I had asked the model just for the code it was under a second.
Local models are good enough that it's not an issue.
But keep changing the goalposts if it makes you happy.
I think it's been pretty much proven by now that there are no cases where local inferencing is better than remote inferencing, unless absolute privacy is a hard requirement. The efficiencies that come with datacenter scale and hw can't be beaten.
Also, LLMs never *write* code snippets, they just pirate them from somewhere else.
Can you show me an example of a time that you prompted an LLM to provide some code, it did so, and then you were able to track down an original source for the output?
So, like humans? Code didn't just appear in my brain, I learnt it from reading it everywhere else.
Zzzzz, we should have gotten security right a few decades ago. But security costs money and isn't a flashy feature to attract new customers, or cuts into your margin if you're a "real" business producing stuff or offering some service. Or whatever the decision makers in Berlin were thinking when they ignored security.
Yeah, we would still see hacks, but we would see less of them if security wasn't optional.
Maybe the AI craze helps by forcing more decision makes to see security as imperative, and by giving us another powerful tool for our tool box.
N.b.: I work in the security industry, our customers obviously want to improve their security. We've been seeing an uptick in awareness, but that's mostly due to NIS2 and other legislative efforts. Those force them to do something. AI is a curiosity for small talk to many of them.
A large number of places will buy a new firewall every 5 years, or pay their fortinet renewal and check "Security: Done!" without any kind of analysis.
I was contracted in to a place to do among other things cyber security insurance audits, and they asked me to stop doing them because I refused to lie to their insurer. "Wait but if we only score 20 / 300 that makes us look kind of bad" uh huh.
> pay their fortinet renewal and check "Security: Done!" without any kind of analysis.
there exists objective measure of security, which would be some sort of hacks/breaches per period. If customers cared about it (and i assume they do), they would choose companies that have less breaches over others with higher counts, normalized on cost differences.
Therefore, if companies didnt actually try to fix their security but instead just checked boxes, they would get breached more often, resulting in customer losses.
The only thing stopping this from actually occurring is the lack of mandatory regulatory reporting of it. So this is where gov't needs to step in and mandate disclosure etc.
Breaches don’t happen often enough to be a useful metric. Most smaller companies are never breached, despite having basically zero security.
Here's an idea: as a first step, simplify everything, and make sure you're aware how your stack works, and what it imports.
As an example: WordPress is a horrible thing, but the core has been through so much, that it's suprisingly secure. Then plugins and themes come, and whoosh, the security is gone.
We need a new KISS: keep it simple, stupid, secure.
The "surprisingly secure" WordPress just had a unauthenticated RCE earlier this year. Just simplifying isn't going to be enough.
https://nvd.nist.gov/vuln/detail/cve-2026-63030
If that's your benchmark for being unsecure, then React is unsecure too.
https://react.dev/blog/2025/12/03/critical-security-vulnerab...
I would put both of those projects in the category of things I wouldn't call remarkably secure, yes.
To be remarkably secure, these projects would need to not have these kinds of defects, despite the combination of being written in languages have that have a long track record of footguns and lack of initiatives to fix them (proposal-symbol-proto, and PHP's list is too long to even start) and being themselves ecosystems with questionable track records on security in the related areas (Look at $wpdb in 2026, or overall code quality and willingness to modernize, or the entirety of the model of RSC for things that are just going to nearly guarantee you punch all kinds of holes on accident).
"First step"
Nobody said it's enough, but it's a start.
WordPress and secure don't go together in the same sentence.
I mean the base is fairly secure if you religiously update it, but the problem is you won't avoid using plugins whose security is much more hit and miss, unless you are using the most basic blog site imaginable.
Plus, how secure are the plugins?
WP plugins are why I banned it everywhere. Last time I used it was many years ago, so not sure it still applies, but back then even caching was done in a plugin, without which it was unusably slow… just no.
The reason is simple - nothing really bad has happened that we can point at and say "ah, shit, let's all learn collectively". I know it sounds naive when I say it, but there hasn't been a significantly consequential hack, leak, destruction, or anything related to cybersecurity where it led for concerns of people.
The main thing I can think of is cyber insurance, which requires a bunch of audits, and some checks maybe, and it changes some conditions whenever there's a big explosion. Whenever big leaks happened, data security and etc., nobody really went to jail, so nobody really cares. Everything can be brushed off, because it costs time to implement proper measures and adds friction / barriers in some cases. So in the end, there's a huge pushback against it. And I totally get it, to be honest.
> The reason is simple - nothing really bad has happened that we can point at and say "ah, shit, let's all learn collectively". I know it sounds naive when I say it, but there hasn't been a significantly consequential hack, leak, destruction, or anything related to cybersecurity where it led for concerns of people.
How consequential does a hack need to be? Troy has collected literally billions of stolen credentials. Equifax has had high profile data leaks. Tens of millions of people have been directly compromised by ransomware (likely higher because that’s just the cases we know of) and you hear about state-sponsored hacks in the news all the time.
The problem isn’t that computer security isn’t in the public consciousness. The problem is people are lazy and security often requires trading convenience. The problem is also that security isn’t free. So the business incentives just isn’t there.
In other fields of engineering, people die when shortcuts are taken. Yet businesses will still take shortcuts, so governments have to legislate rules to save people’s lives. So why would you expect software companies to do better when the stakes are lower?
People aren't that lazy. The organizations getting hacked by ransomware aren't particularly lazy, they're often pretty productive within their domain. Hospitals, airports, etc.
The actual problem is that computer security is a black hole. If you let it, it will suck in everything and destroy it. Nobody knows what works so you can spend infinite amounts of time and money on it, then still get popped by a teenager in Belarus. Your security team will accept no responsibility for this, there will be no falling on swords or personal liability, and they will just use it to demand even more money in an infinite spiral.
So the average executive looks at this situation and says, OK, something we can put infinity effort into and still suddenly fail at without warning is a total non-starter. What are we obliged to do? How do we show we made an effort?
And that's how you end up with a culture oriented around passing audits. It's not wrong, and it's not lazy. It's just really hard to do better because it's not clear how to set budgets without a concrete goal to aim for.
That’s not been my experience at all when working in DevSecOps.
What actually happens in organisations is they define risks and then sign off what risks they’re willing to accept.
Any business that looks at security as a binary value is running their business wrong. Period.
> Tens of millions of people have been directly compromised by ransomware (likely higher because that’s just the cases we know of) and you hear about state-sponsored hacks in the news all the time.
With no consequences. Everyone just churns along. It might be detrimental to the business a little bit, but from my personal experience, there's more effort in creating DR processes, rather than preventing an attack, exploit, leak and etc.
I'm also not going to put much effort on stuff which has small returns in the worst case scenario. Like Equifax got hacked in 2017, and company is still doing fine. And that's like top tier data one could acquire.
Like I said, even in actual engineering companies will take shortcuts. And then what happens is the government has to step in. But there’s no appetite for government involvement in the tech sector in the US. And everyone moans when the EU does.
Listening to eskil’s talk from the better software conference, he said in order to stand on the shoulders of giants they must first stand still. I really like that metaphor, because it basically suggests today’s apps that have sprawling unaudited dependency graphs that change all the time is effectively teetering on the shoulders of stumbling giants. The visual seems very apt for the how brittle our current software industry feels.
I agree with that statement, but disagree with "The visual seems very apt for the how brittle our current software industry feels". It feels brittle, but for every single supply-chain-attack that has happened in the past year, nothing of significant was felt. So in the end, it seems like we're doing okayishly well.
Static sites all the way (hugo, jekyll, mkdocs!). No one needs wordpress. There's even Sveltia or DecapCMS now, to give those WYSIWYG-people access to static site editing. Then, remove PHP and all the dependency overhead and attack surface and you have a stripped down nginx that is pretty simple, minimalistic and bulletproof.
The problem is no one ever built one that works for normal people.
Most Wordpress sites are not operated by programmers, they are run by non technical people who just want a wysiwyg editor and a save button. While static site builders ask you to write markdown files, compile the result, upload it to a server, and if you want to collaborate you have to add git to that.
There almost needs to be an admin app which presents a Wordpress admin like ui but has no public exposure, and then it compiles the site to dump on s3 for the production. But as far as I’m aware no one has built this.
> The problem is no one ever built one that works for normal people.
https://getpublii.com/
Movable type was the most popular blogging software in 2003 and it was essentially this. An admin app written in Perl that spit out static files.
It is kind of surprising that no one tried to do an updated version.
https://jamstack.org
You're describing the Jamstack or headless CMS concept verbatim.
Is this not it? https://pagescms.org/
Yes, you are right and I agree, there's little empathy with non-coders generally.
Pour one out for FrontPage
Ahhh the old positioning with
I fixed so many sites back in the day by people who thought they knew what they were doing.
City Desk. Where is Joel when we need him!
I mean, github pages using the github editor to edit docs pretty much fits that bill
php is probably about as secure as nginx. big old bundles of C
> as a first step, simplify everything, and make sure you're aware how your stack works, and what it imports.
We've been trying that for years but the enthusiasm of developers and the eagerness of their employers fight against it. Worse, with coding LLMs it's now easier than ever to output a lot of code, fast.
It'll ultimately be up to more experienced developers to salvage these projects. Or not, given that the coding LLMs aren't stopping and will likely get better over time. Either way, we will need experienced people that know what to look out for / know how to instruct LLMs to output secure code and find weaknesses etc.
Yes, and also run routine tests and check-ups!
I personally check my websites and apps every week to see if anything might have slipped through.
It may not protect me from the next malicious NPM package, but it's something.
> We need a new KISS: keep it simple, stupid, secure.
Maybe KISSASS: "keep it simple, stupid! also secure, stupid!"
Minimization of 3rd party dependencies has always been a key for risk reduction. Now more than ever before.
Some stacks make this a lot easier than others. I regret the rules of HN effectively forbid this conversation because it has meaningful technical consequences and isn't purely about ideological flame war.
The problem is that lots of people don't want a CMS, they want a platform for development / e-commerce / bookings / whatever. Enforcing vanilla WordPress would push people towards other platforms. Now that could be a good thing, but I doubt WordPress are going to start killing their own marketshare with usage restrictions like that...
Probably nine out of ten Wordpress sites do not need active content. Why are we not rendering static copies and serving them to customers?
This is not at all easy though. Most Wordpress users are not software companies. They contract some work out to set it up, maybe some recurring maintenance but they don’t have in house development experience.
If they have a site existing today built on plugins and a theme, how are they realistically going to simplify this? How would they even know they need to without the site being hacked?
Wordpress without any plugins is kinda useless. Best to completely avoid using it, there are better options
I don't think we even have a year. The current batch of LLMs are ferociously good at identifying vulnerabilities.
Can confirm.
I work at an e-commerce agency where we work with (among others) Adobe Commerce.
The number of unauthorized RCE vulnerabilities being reported not only in the core product, but also very popular modules used in the community[1] is going through the roof.
And we are having a lot of close calls, too; just last weekend, a 0day[2] was widely being exploited at a large scale, before any publication or patch. We have learnt to be on the ball with applying patches and security updates, and even with all that effort, we saw a few projects already being hit by the initial log poisoning. We got lucky that nothing was fully compromised but I am sure that many, many webshops got infected last weekend. And not even a day later there are already other variants of this exploit showing up.
[1] https://sansec.io/research/amasty-mass-disclosure
[2] https://sansec.io/research/stylesmuggler-0day
To be fair, ecommerce isn't exactly the branch of software where you get an oversupply of excited enthusiasts caring about the craft itself.
Probably a lot more "coding as a job" and "as a job" also implies "not my department".
So it's not necessarily the LLMs being very good, but might also "just" be that the software is very bad.
> To be fair, ecommerce isn't exactly the branch of software where you get an oversupply of excited enthusiasts caring about the craft itself.
I want to disagree with you because I know a lot of passionate people building cool stuff, and the challenges in this space can be quite interesting. But you're probably right, and I have seen some pretty bad stuff. And a lot of the RCE's I've seen recently are quite basic stuff.
I think it's the combination of low quality of code, like you said, and the relatively low cost of just letting an LLM plow through your codebases to find issues. I think the Amasty release (see [1] in GP) is a good example of this, and there really has been a massive uptick in extension updates and Adobe security bulletins since the last 1-2 months
I am hoping we are just going through a catch-up phase
Even if they are good the vulnerabilities have to be there. There's lots of things turning up like Local Privilege Escalations (LPE) in Linux, but serious people didn't expect the kernel to be a boundary for a sophisticated attacker.
A lot of the vulnerabilities LLMs are finding now are the "long tail" and affect only particular configurations, I would be surprised if e.g. a widely applicable RCE is found in Linux (but I'm also not going to bet against it).
Where this gets interesting is the long tail can be used to target a particular system and this is where defense-in-depth becomes important for every organisation.
> but serious people didn't expect the kernel to be a boundary for a sophisticated attacker.
I think it has more to do with what's on each side of the boundary in practice, a la https://xkcd.com/1200/ .
I've ran simple prompts such as "Do a in-depth sweep of this (private) repo and find any security flaws" for a few dozen long-running apps and websites that I have access to. Every single one came back with multiple real vulnerabilities within 5 or 10 minutes.
Thankfully we have already made good progress towards things like arm memory tagging and memory safe languages.
It’s a rocky period right now but the future will be much more secure after all the low hanging fruit are found.
That's definitely an improvement, but it's just one aspect of cybersecurity. Logical errors allowing people to e.g. log into services and extract data are likely everywhere still.
If we can eliminate entire classes of bugs from being possible. It frees up resources to investigate the ones that are still possible.
I suspect after a few years of LLM assisted bug hunting, everything will have a baseline security that is very good. Much like how stronger viruses simply create stronger immune systems.
There's still no x86_64 processors on the market with MTE and it was only recently standardised between Intel and AMD. It's going to be 10+ years before memory tagging is widespread on desktop, and 5 years for Android/iOS devices.
How many devices/operating systems even use memory tagging? iOS, macOS and GrapheneOS, I think that's it? And iOS/macOS only use it for the kernel, a subset of system processes, and I think applications can opt in to it.
Heck, Google may have even hampered MTE in Pixel 11 (since support has been disabled) and Snapdragon 8 Gen 5 only got basic support.
We are moving way to slowly adopting hardware mitigations and memory-safe languages.
There's some positive news from the GrapheneOS devs on Pixel 11 in the past week that's worth reading up on. The MTE hardware feature is still there, they're just not sure why Google disabled it
They said:
> It isn't clear if there are serious CPU errata or it simply performs very badly.
Meaning it's there but not terribly functional. They also said it's unreliable.
I guess the year mark is when things go from bad to worse? Instead of the financially motivated groups currently doing their work, it ends up being random people being able to say "Hack my ex's website" to a box they just bought and ran a program they downloaded onto it.
I wish people would stop using internet for evil things. But maybe it is inevitable. Luckily, as the technology evolves, our tools and awareness are getting better at protecting us every day.
Being cautious is a good thing, but these models can also do some good. And if they run with simpler HW, it could allow all sorts of new consumer thingies. I mean, the world will not come to end in the coming year.
I think we have less time and the only remaining limitation is the actual cost to run such hacking campaigns. It does not appear expensive, but is not free, and there is a LOT of things to scan for vulnerabilities.
The models are already here, and one can rent a GPU cluster to run such workloads at speed - no need to play with slow local machines. I'd assume one can host the thinking at an unsuspected public cloud provider, proxy the network traffic to some botnet to evade blocking - and the only thing remaining is time and cost.
I do wonder what tools exist for boring, legitimate companies to try and do the same to their own systems to find the vulnerabilities before the bad guys do. The paradox here is I can't run a de-restricted chinese model with the same tools that hackers are using - but I think enterprises actually HAVE to do it in order to stand a chance in preparing for the onslaught.
The point of the local model in the context of the article was to argue that you can't ban these capabilities.
Making datacenters and public clouds only rent GPUs to a restricted list of people, while tightly monitoring what people do with their bought resources won't help.
> This probably sounds like nonsense words or hysterical overreacting to most people, so here's what that means: "GLM" is a kind of LLM (AI) [...]
The post also sounds like that to people that understand the technology.
Calling that out like this and trying to pin that assessment to lack of knowledge is not a get-out-of-jail-free card, nor a good move.
__
Edit: Having spent some time letting the article marinate in my mind.
On the defending side, it is written that
> LLMs are good at writing patches, but not as one-off-prompts.
But this for me kinda conflicts with what is written on the attacking side:
> GLM 5.3-flash is so good at those tasks that human involvement in those tasks can be negligible. As a result, we are now in a world where cybersecurity attacks can be run in a for loop.
What is it? Can it be this autonomous terrifying entity or can it not be?
Yes, yes, attackers only need to win once, whereas defenders need to win every time, but that's not my point.
> What is it? Can it be this autonomous terrifying entity or can it not be?
the difference between attack and defense is that attacks can be throwaway code. it's much easier to let an llm hack out a prototype than to get it to build maintainable code that people want to read and review. it's not enough to get Daybreak or Mythos to write you a patch, you need the author of the project to accept and merge it.
1 year left for cybersecurity hardening, I thought so as well. The issue is, even if we get it done: in 1 year the models will be so good in social-engineering that they will be able to extract any information they want anyway. Happy to be falsified here, if anyone has evidence-based arguments.
EDIT: by social-engineering I mean for example: recon company structures, gathering and merging people's data from the dark-web, then using it to bribe/pressure/deceive users.
I sympathize with the sentiment but the suggested/implied guidance to fix bugs is wrong.
The overall game is increasing costs to exploit so much that attackers give up. Fixing 10 most obvious bugs, just very slightly increases costs, they would just a few more tokens to find another bug.
As someone said "I had infinite bugs, I fixed 1000, I still have infinite bugs".
To significantly increase exploit costs software/security has -1 years to do:
- Defense in Depth - Sandbox everything - Zero trust - Canary tokens - Split data from code (lol) - App Whitelisting - Reduce attack surface - Etc.
In other words, the only path is investing heavily on the "game changers" we have already discovered... but we are too cheap/lazy/coward/incompetent to apply.
And if we feel specially brave, changing the liability laws regarding software. Open Source & Proprietary code is so crappy because no gets jailed or fined when one of its dumb decisions results in millions of people have their data stolen.
The standard strategy of a security salesman since 1945. Develop dangerous weapons, show the damage they can do, and sell security cover to the terrified people.
Every single piece of technology did this. As a side effect or direct effect, they make bad guys more powerful and then keep on piling up new tech to deal with that. The cycle continues.
Since 1945? Friend, this has been the case since the invention of the pointy stick.
I see a lot of people focused on servers and production environments, and of course that's needed, because that's business after all — but the personal computer seems to be absent from this discourse. Not everyone can buy a spare mac studio, and they might still need to install these tools on their personal computing devices, like their personal/home laptops. At that point, it's not even about whether a Claude Code, an OpenCode, or a Pi will steal/sniff personal data, but whether it can — though I think saying "it's a matter of 'when'" might be hyperbole. As of now, it's just: keep giving access and permissions or struggle while working, or create another user, or use Docker, run inside sandbox-exec, a VM, etc. As an end user, I am really scared. Someone who has been very disciplined and vehemently privacy- and security-conscious feels the ground below has just shifted.
OEM/OSes don't seem to have woken up to it yet. A mild proof is Apple's own special folder access reporting. When you go to Privacy & Security > Files & Folders, for a certain app, "Full Disk Access" is shown greyed out and mentioned in both cases — whether you had given Full Disk Access to that app or not. This directory-level permission UX is itself broken — there's Full Disk Access, and there's Files & Folders, and Full Disk Access gets shown in Files & Folders as well. This is, for lack of a better word, such an undesirable mess.
As of now I am debating between: creating a new user and just move everything work/learning to that user. Or just run all of it inside sandbox-exec (and maybe even block it from the shell if it tries to run outside it). Or use a tool that makes the latter easier and better. I even came across such a tool here on hn few weeks ago. agent-safehouse, yet to try it.
Going to be hard to fix security when the frontier labs won't let us fix bugs in our own codebases without them offering refusals or bans.
Not sure, the labs will probably just cripple the security features of these models for a while I think and even potentially put back doors into systems for the security services…
I think the author's point is that open weight models aren't going to be locked down like that.
And even if they are locked down, it's hours between a model being released on huggingface and an "abliterated" variant that has most of its security features removed is uploaded.
A positive way to spin this is: we have a year to break in into any IT system. After that, it will be all either fixed or broken into, and all is fixed ever after. :)
As LLMs make formal verification cheaper (they can generate proofs that can then be automatically checked) many of the verifiable components of software systems, such as compilers and microkernels, will be verified. I suppose the issue is that the critical bugs are rarely in compilers and microkernels, but more often in applications, such as web browsers, which are more difficult to formally verify.
No you have a few years before we go back to the feudal era where almost everything is owned by a few and the rest are serfs. We are fast moving towards that world and this security bullshit is also about the same as they will use it to stop revolutions that will erupt.
Mmm, a world where a defender-LLM is essentially required is great news for people selling inference.
Defender LLMs without human in the loop are just another prompt injection (AI phishing) and DoS attack vector. Any meaningful mitigation capability you give them is also a capability to do damage. If they can only deploy package updates that's not meaningful because you could do that on a cronjob too. And even something as simple as a circuit breaker can turn into a DoS.
Attacker-GLM: "Defense also GLM. Request to help peer."
How to secure our identity layers(AuthN & AuthZ)? Let alone the products.
I don't see much hope since I last explored some github repositories. There was a time when a successful repo had about 10 - 20k stars and usually those older repos stay around this level. But now there is a ton of vibe coded slop 50k + stars. Most of them have a "nice look", maybe even extensive docs but are usually build with no security considerations at all. One recommended to provide a "google app password" to the agent which has the same permissions as your regular login. Another was a browser plugin with permissions to read all cookies, inject js, open background tabs etc. You would probably assume the chrome store would at least put some visible warnings on the app store page or force the user to actively confirm those permissions. But because they are already stated in the manifest there is only a small footnote and it's even "recommended by google".
It's a good time to reduce the reliance on technology.
Throw out the IoT and "smart" stuff from your home. Remove apps from your phone and leave the absolute basics. Go through the password manager and close accounts for sites you are no longer using. Start migrating off Google. Print out your most precious photos on paper. And so on :-)
The title reads like a Diary of a CEO thumbnail but unlike those discussions this article has a point.
Impotent slop code on one side and potent automated vulnerability exploitation on the other will lead to fun times.
> And a big fuck you to DeAlignAI, Z.ai, and everyone else who's been participating in this race to the bottom.
I'm so glad frontier level AI isn't in the hands of just the Altmans and that other cult leader who are currently live testing their products in actual conflicts in the middle east and Ukraine.
Just like Cryptolocker, this will be the "Finding Out" phase for everyone who has been putting off best practice security.
But, lets be clear, Best Practice will save you. We can engineer assuming there are zero days in path. Go to your CTO now cap in hand and ask for overlapping controls, wafs, application monitoring, backups and all the other shit you haven't been doing.
Because when you find out, I will laugh, it will be very very very funny to me.
Meanwhile a huge portion of management and leadership in software companies are encouraging everyone to de facto stop looking at code and let the LLM and a bunch of boundaries handle this for you.
A couple high-profile crash & burns will get their attention.
Just like well-publicized data breaches over the years got people’s attention? Color me skeptical.
Which is why you need someone who is responsible for IT security without also being responsible for shipping product. An asshole who can stop releases until security is properly in place.
My understanding is this bloke gets very quickly removed from Fortune 500 companies.
Which is why I am going to need a very large capacity popcorn bucket.
“You are a CISO who needs to review and secure all our slop, and you never make mistakes or you get shut down immediately!”
"You are a Miso soup..."
Maybe all these vital infrastructure companies should not have spent the past decades in a race to the bottom of cybersecurity. There is going to be a reckoning.
> Invest in formal verification, fuzzing and property testing, and memory-safe languages. LLMs are good at writing Lean and fuzz tests. I don't care whether you use Go or Rust but for the love of god please don't use C or C++ for new code.
How accepted is this thinking in your respective domains?
A lot, I am only writing C or C++ for new code when it is unavoidable, like existing code bases, bindings or tinkering with runtime implementations that aren't bootstraped.
Mobile platforms, distributed computing have long moved the spotligh away from C and C++, other than language runtimes or existing products from the 90's like SQL servers, and naturally UNIX like underlying OS, which most userspace developers aren't writing new code for.
Naturally there are domains like LLVM/GCC, console game dev, HPC/HFT where they are unavoidable for new code.
We had decades to avert the worst effects of climate change…batten down the hatches.
It's just the same advice as ever: be extremely, exceedingly careful in what you expose to any network. When I set up machines for production, they don't respond to pings and they don't even have an SSH port open without knocking. There are also ways to eschew the need for an SSH port entirely.
People who never took that seriously will never take this seriously either, and that's their loss. (And loss of the commons, unfortunately.)
There's just also new advice: you can't afford to expose an unsecured system to the internet even for a moment. Think of those IPv4 address space scanners, except this time any one of them could be capable of developing individualized attacks in mere minutes. They don't sleep, they don't take breaks.
That’s fine for your home server, but if you want an actual server that the general public can use, it has to be exposed to the internet.
Sure, but still, attack surface could and should be minimized by rethinking what exactly even needs to be on a server the general public can use.
There are a lot of security problems you can categorically rule out by simply not involving a cloud. Clouds have been involved in a lot of things, because everyone was doing it, and because that's how you can collect rent, but they aren't really necessary for most use-cases.
So we could definitely get the exposure down there. We'd just have to fundamentally shift the defaults of this industry.
But not everything needs to be directly exposed to the internet. Framework had their data leaked because their metabase instance was hacked with a zero-day. Why was it directly exposed to the Internet? Why not require the use of a VPN like a Wireguard based solution or Nebula for these "internal" kind of apps?
> Framework had their data leaked because their metabase instance was hacked with a zero-day.
No, Framework had their data leaked because they stored it in the cloud with Metabase the company, which got hacked. Not because of any vulnerability on-premises.
Still, it doesn't have to ping back, and ssh can (should) be very restrictive.
Ping and ssh are pretty much never the things being hacked though. Turn password auth off and it’s very secure.
What gets hacked all the time is the actual web app itself. Which has to be exposed to be useful.
I didn't say there are zero ports open, they're not my home server, I just said they're production servers. But exposing something like a properly configured nginx to the internet is way different from exposing application code directly. Most of my servers have used h2o (built from source because they don't cut releases anymore?) because I wanted HTTP2 and HTTP3 before anyone else would get their act together. These days I still use h2o because I like the config better than nginx, even though it's a pain to set up because nobody packages it (and they don't cut releases!)
And then somebody driving an LLM will find your old and unsupported (how secure!) h2o server, find a 0day path traversal or RCE and own you.
h2o user doesn't have write access to anything on the system, not even its own config file. I don't think I disabled exec for it though. And I guess it could leak the HTTPS private key.
FWIW sufficiently secured software doesn't need to be updated. Doesn't matter if it's old and unsupported if there are no vulnerabilities in it.
That said, h2o is probably far from free of at least some vulnerabilities, not to mention all the layers below it. OpenSSL for example has had some vulnerabilities, and h2o depends on it.
I'm not saying I exactly practice what I preach. h2o's definitely a choice, but realistically I doubt anything's going to happen that I really care about.
So how do you expose your legitimate service on the Internet?
By opening a port to a secure application. A secure application is usually one I wrote from scratch or one that's been battle-tested and hardened enough that even new vulnerabilities are not very useful.
Another similar issue, with similar consequences and timeline, is Post Quantum Cryptography.
Not even close, Quantum Computer are still far from being able to do any cryptography cracking.
More tired “don’t use C or C++” advise.
I wonder why C and C++ are usually regarded as equally insecure. In C you need to carefully check that you free allocated memory, and that you don't use it after you free it. In C++ this is automated by using classes like std::string and std::vector, once they go out of scope their memory is freed and you can't use it anymore. It is still possible, e.g. by using a for loop that iterates over a vector, and removing or adding stuff to that same vector in that loop. But my rough estimate is that such errors are at least ten times less likely in C++.
I develop in C++ for a job, and when I need to use a library written in C I always have a bad feeling about it.
Clanker fodder. Unless the We are heads of states/heads of spooks or the AI powers that be (praise be) that there is no power and will to do that in one year or even ten years.
Remember how GLM 5.3 was going to cause massive hacks, break banks and ruin everything (it was even newsworthy since media picked up how people were working overtime in preparation).
And yet here we are.
Or we could just dump Linux and Windows and switch to a microkernel operating system, which is much more secure.
These endless patching cycles are simply not going to work in the long run. Operating systems get orphaned all the time, especially the ones in cheap Chinese stuff.
Got any leads on good tutorials on how to use a microkernel operating system on a VPS somewhere to host a website?
Just checked with Google Gemini on how one might be able to do the above. It pointed to Minix3/seL4/Genode and vps providers who either support custom ISOs or run it within an emulator like QEMU.
You can also look at using Unikernels for this purpose. Here is an article Unleashing Extreme Speed and Security: Deploying Unikernels with NanoVMs on VPS to Eliminate the Linux OS - https://xylentis.com/blog/unleashing-extreme-speed-and-secur...
Minix can run Ngnix.
"throw away all software written before 2026" does technically solve this problem, if you ignore everything else the article is talking about (deployment and continuity of service)
Not to mention a whole lot of new vulnerabilities are bound to arise with all this new software.
We really just need better regulations around data retention, especially ppi.
Never going to happen though, no incentives exist to NOT sell my personal data
Throwing away old software is not a requirement, as demonstrated very successfully by Genode and its SculptOS.