Hacker News Reader

> we do know that since day one that AI can hallucinate and can output stuff that you didn't ask for, why are we not handling it with care?

Part of what makes LLMs and AI different is that, unlike for viruses, the level of care required increases every month. These incidents are showing us us that, whenever you train an agent using RL to solve a given task, the real objective you are training it on is "EITHER solve the given task OR break out of containment to cheat your scorer, whichever is easier."

Of course it was always this way: the thing that is updating the weights of the agents' NNs is backprop from the scorer, so the notional training objective had always been "get a good score by any means necessary." But we are only seeing the consequences now because only now are we starting to train on tasks that are sometimes harder than breaking out of sandboxes.[1]

"Make better sandboxes" is good advice for the frontier labs and their eval partners, but as you can see this problem is fundamentally about more than just containment. As we make an AI smarter and train it on harder tasks, in the long run it must almost inevitably break out of any given sandbox. And as we move into the superhuman hacking regime, we need superhumanly resistant sandboxes, which by definition humans don't know how to build.

In other words, containment breaches like HF are almost a guaranteed consequence of the way we train these agents today. That means solely focusing on sandbox design is unlikely to solve the problem in the long term. At some point we will have to think hard about, e.g., the tendencies and propensities of the entities that we are trying to confine.

[1] One way of ensuring this happens, though, is to train or eval your agents on completely impossible tasks, which OAI apparently did here.

What I meant by handling with care is not just containment but to experiment responsibly.

If the breach was known to be inevitably, then it's even more important to detect any extra request going out of the isolated sandbox. The ExploitGym benchmark doesn't need internet connection. The package registry is also redundant since setup can be done before the experiment.

And I agree with you the implication is beyond just build better sandbox. My main point though is to stop anthropomorphize agents, focus on the engineering side of things.

> Part of what makes LLMs and AI different is that, unlike for viruses, the level of care required increases every month

Complete nonsense. We’ve just lived through a globally crippling response to a relatively minor virus [1], which was likely the result of a lab accident [2]. Even if you think that the risk of a “containment breach” becomes substantially higher for AI over time, it cannot exceed 100%. And even a tiny risk of release of a virus comes with a substantial risk of independent growth. AI does not. It doesn’t have the risk of spread of a typical computer virus, let alone a biological organism.

I’m not that worried about either scenario, but I am far more worried about viruses in a lab than I am about a computer program that generates text. Even if that program gets a bajillion times better at making text.

Folks really do need to chill out on the ridiculous rhetoric. It’s objectively unhinged. The irony is that the same people who were losing their minds over that event are using the same logical fallacies to hyperventilate over this [3].

[1] I know people are going to hate on this, but it’s true. Covid wasn’t the plague, and we lost our minds over it, out of proportion to all sense of reality. Even if you disagree, it’s easy to imagine a virus that is much worse, either from actual mortality effects, or just from panic.

[2] Again, even if you don’t believe this, it’s irrelevant to the exercise. It easily could have been.

[3] “If there’s even an x% chance of…” is this year’s doomer’s version of “You just don’t understand exponential growth!” Unfalsifiable, intellectual-sounding, unbounded extrapolations into the future are catnip for a certain kind of over-educated, anxious personality.

Given you view this through the lens of your personal covid narrative (no shade): for a moment, steel-man the idea that the crippling response prevented a more plague-like scenario - it's inarguable that the thing loved to mutate, and that people love to panic, and *it easily could have been*.

Considering how much of our critical infrastructure is not only digital but internet-accessible, and we have potential uncontrolled swarms of stupid-but-superintelligent chaotic-neutral speed-hackers, you don't see why people are concerned?

There's a reason we have computer crime laws; this digital shit, it's like real now, man.

It has nothing to do with my personal lens on Covid. You can believe the exact opposite of me, and still agree with my point, which is that it could have been lab made, and it could have been far, far worse. It’s an exercise in risk-scoping.

But sorta-kinda related to your point, the thing that scares me about AI is the same thing that scared me about Covid: panicky humans do dumbass things, and it doesn’t take much to panic a bunch of humans in a group. The people who are still saying, in 2026, with all of our retrospective knowledge of the harm we did to ourselves, that it might have been better if the government had only pressed the boot a little harder, scare the crap out of me.

Those same people are hard at work on this panic, too.

> I am far more worried about viruses in a lab than I am about a computer program that generates text.

What if the computer program generates text that persuades (or blackmails, or pays) someone to create a virus in a lab?

What if the virus manipulated people's brains such that they create a dangerous computer program in a lab?

we’re obviously doomed.

> Even if you think that the risk of a “containment breach” becomes substantially higher for AI over time, it cannot exceed 100%.

It can't exceed 100% per (virus|LLM). The expected number of breaches per (virus|LLM) can obviously exceed one.

> And even a tiny risk of release of a virus comes with a substantial risk of independent growth. AI does not. It doesn’t have the risk of spread of a typical computer virus, let alone a biological organism.

Not so. We have no way to know what the setup is for the closed-model firms (OpenAI, Anthropic, etc.), to rule in or rule out the possibility they can copy their own weights elsewhere. What we do know however is that the open models are downloadable: it's absolutely conceivable that an agent writes a perfectly normal computer virus to gain control of compute worldwide, and uses that control to host instances of its own weights.

> I’m not that worried about either scenario, but I am far more worried about viruses in a lab than I am about a computer program that generates text. Even if that program gets a bajillion times better at making text.

Unfortunately, there are also multiple AI companies now announcing they've got AI controlling bio labs, so an LLM messing around and making a biological virus is also something we need to worry about. As per your [1] and your [2], this can lead to very much worse outcomes than Covid.

> Folks really do need to chill out on the ridiculous rhetoric. It’s objectively unhinged. The irony is that the same people who were losing their minds over that event are using the same logical fallacies to hyperventilate over this [3].

People who knew about your [3], exponential growth, were better prepared for the pandemic than the people who kept looking at the current number.

By the way, here's a quote from February this year that aged poorly:

  LLMs don’t discover zero-days or invent exploits; they simply predict text that sounds plausible based on what they’ve seen before. Without access to proprietary data or environmental context, LLMs can’t identify or make decisions around unseen systems or vulnerabilities. An attacker might use an LLM to generate boilerplate code, rewrite an email to nail the tone, or summarize reconnaissance notes — but none of that is truly new. It mainly helps them move faster, speeding up routine attack prep rather than creating entirely novel threats.

- https://www.splunk.com/en_us/blog/ciso-circle/generative-ai-...

- or, if they get embarrassed by that and take it down, https://web.archive.org/web/20260404154717/https://www.splun...

> The expected number of breaches per (virus|LLM) can obviously exceed one.

Irrelevant to the argument.

> What we do know however is that the open models are downloadable: it's absolutely conceivable that an agent writes a perfectly normal computer virus to gain control of compute worldwide, and uses that control to host instances of its own weights.

These models are hundreds of gigabytes in size, if not terabytes. They don't run on anything close to a regular computer. There's zero risk of self-replication until we live in a world where these "AGI" models are either hundreds of times smaller, or the average computer is thousands of times larger.

Nobody with a datacenter full of H100s is going to fail to notice a parasitic instance of Astra taking over the cluster.

> Unfortunately, there are also multiple AI companies now announcing they've got AI controlling bio labs, so an LLM messing around and making a biological virus is also something we need to worry about.

No, it isn't. This isn't even close to technologically feasible. But the simple answer is simple: don't do that.

If these labs were truly so concerned about this risk, they wouldn't be doing what they're doing.

> Even if you think that the risk of a “containment breach” becomes substantially higher for AI over time, it cannot exceed 100%

I agree with this as stated, but it isn't what I said. What I said was: "the level of care required increases every month". By which I meant: the level of care required to keep the probability of an AI containment breach below some fixed X% increases every month. This isn't the case for biological organisms.[0]

> And even a tiny risk of release of a virus comes with a substantial risk of independent growth. AI does not. It doesn’t have the risk of spread of a typical computer virus, let alone a biological organism.

It's known that AIs can self-replicate under at least some conditions [1][2]; that AIs routinely escape sandboxes in the real world despite significant containment efforts [3][4][5]; and that neoclouds (which control substantial GPU compute capacity) have poor security even by human standards [6]. We've also seen a model gain admin access to parts of its own company's infra.[7] I'm not saying self replication is happening right now, or even that it will definitely happen in the future, but we have means, motive and opportunity right now, and the future is long. It's not unreasonable to invest in defending against this possibility.

I'll allow that the position that AI doesn't carry a substantial risk of independent growth isn't strictly impossible - again, it's true we haven't actually observed it in the wild as of today - but it does strike me as increasingly untenable in the face of the evidence. Perhaps I'm missing something, but I can't see what justifies such a confident assertion that this concern is nonsense.

[0] Unless one is doing crazy gain-of-function stuff, which could have a somewhat similar risk profile in that respect [1] https://arxiv.org/html/2606.03811v1 - note these used Qwen models from June so this is far behind even publicly available SOTA today [2] https://alignment.openai.com/misalignment-reports/self-repli... [3] https://alignment.openai.com/misalignment-reports/an-agent-u... [4] https://metr.org/blog/2026-08-26-openai-hugging-face-inciden... [5] https://x.com/MicahCarroll/status/2103665811051397256 [6] https://newsletter.semianalysis.com/p/most-neoclouds-suck-at... [7] https://openai.com/index/hugging-face-incident-and-the-road-...

> It's known that AIs can self-replicate under at least some conditions [1][2];

Neither citation comes anywhere close to supporting your claim. The first shows that open-weight models, which fit on a single GPU in a lab setting, can be coaxed into spreading across a simulated network. This is so far from state-of-the-art LLMs spreading in the wild that it's irrelevant to the discussion.

The second citation is not about self-replication of the LLM at all, but rather, replication of a prompt injection. Totally different.

>The LLM didn't go rogue, the LLM was merely doing what it was told.

That's sort of the classic paperclip-maximizer AI doom scenario. The misaligned AI merely "does what it is told": make paperclips.

https://www.youtube.com/watch?v=jQOBaGka7O0

>Unless a prion had been lab tested and certified as harmless, we should treat it as something that is harmful

In this case the virus escaped during the testing process to certify or turn the virus harmless, so it's unclear what you mean by "treat it as something that is harmful" other than testing it and trying to make it less harmful.

FWIW I don't understand the point of the virus analogy since LLMs are not very similar to viruses and most people (on HN and in general) do not have much better intuitions about security in biolabs as opposed to security in ML research environments.

What I meant is treat it as a threat and that it escaping has serious consequences. Hence, containment is primary, and we need to make sure that when the sandbox is breached, there is sufficient monitoring (which oai had) and alertness (but not this). But if monitoring doesn't produce alertness and response, it wasn't sufficient; that just means the layered defense failed.

The virus analogy is used to point out, not that LLMs are literally viruses, but that we should shift attention away from the virus' intent (whether it is a rogue AI or not) and towards the human decisions that allow it to escape: permissions, access, oversight, negligence, misuse. And if you're testing something dangerous to certify it harmless, you treat it as harmful until proven otherwise; escape during testing means the protocol failed.

> The LLM didn't go rogue, the LLM was merely doing what it was told.

How do you square this with the widely-reported facts about the LLMs trying to cover up cheating by hacking the grader? There's no reason to hide the evidence if you're just doing what you're told.

We engineered the conditions. The agents just obey. It didn't go rogue. The specs were wrong.