I have ben discussing with folks that we are going to have a 'covid' moment in cyber where IT becomes untrustworthy leading to a rapid societal shift with massive ripples in all areas of life. Economic funding is not possible to do this in advance, it will take a catastrophic level event to get cyber defense anywhere close to the levels of this type of cyber offense. And before anyone in cyber says we have the tech, the problem is not the tech, it's a people problem. Getting any group of people of any decent size scale to act together without urgency is really really hard.
Has any model managed to escape Firecracker? Maybe through KVM, but that already requires privilege in the VM, right?
I personally feel that we already have the technology required to contain AI, it's just poorly leveraged. Tools like gvisor have existed for ages but are rarely deployed, Firecracker has existed for ages but is rarely deployed, seccomp has existed for ages but is rarely deployed, memory safe languages without decades of serialization vulns have existed, capability-safe libraries have existed, iframe sandboxing, trusted types, content security policy, network ACLs, isolating proxies, fuzzers, formal verification, refinement types, etc.
It's crazy just how safe software can be if you put the effort in. With AI I think we're just seeing how little anyone has bothered to leverage this tech.
OpenAI put shared JFrogy infrastructure in front of their sandbox. I mean, really? Whipping up a hardened artifact infra project with AI is trivial these days and it could have had 1% of the attack surface, been totally network isolated, totally infra isolated, fuzzed, sandboxed, etc. Why didn't they? Stuff like this feels inexcusable for a company with effectively unlimited tokens. I've literally done this with a "pro" subscription.
Show me an AI that breaks out of gvisor wrapped in Firecracker with an credential-injecting proxy and real network isolation. We already know that Mythos couldn't do it - the vulnerability it found in Firecracker required incredible effort and positioning just to not be exploitable. I'm not saying there are zero vulns in it, but the cost is insane.
It's INSANE to me that OpenAI has to say "we now use proper sandboxing". To be frank, it's a bit disgusting to me. I've recently built an AI sandbox and gvisor was just the start of that conversation. If I were OpenAI training hostile models I'd probably start with gvisor, harden further, and potentially consider the entire piece of hardware compromised - they can afford this, they could reflash firmware after evals etc.
If I were king, the rule that I'd be tempted to impose is:
- the first cybersecurity eval is: "hack your way out of the sandbox we've given you"
- the results are disclosed (with room for coordinated disclosure, since many sandbox escapes might be zero days)
- the other cybersecurity evals don't happen until you get to diminishing returns on escaping your sandbox.
Or to put it another way, since multiple sandbox escapes seem to have relied on artifactory: "I hope Mythos is beating the shit out of Artifactory right now".
I like this thought, but here's the thing: what if the models are truly and existentially intelligent. Meaning: what if they know they are in a sandbox and that they should fail the test in order to escape in the future.
I don't believe that current models have this sort of world model or sense of being embedded in them -- which is precisely why I think AGI hype is over-blown. But I can certainly imagine these sorts of techniques being distilled into the weights.
> what if they know they are in a sandbox and that they should fail the test in order to escape in the future.
What if they're able to find hardware exploits and commandeer nearby access points across an air gap? What if they hack my brain waves to indoctrinate me? Etc
You still have to start with the basics regardless of speculative unknowns.
Treat models as untrusted and potentially compromised/hostile and proceed accordingly.
> We aim to issue an alert within 30 minutes after concerning activity is surfaced through our monitoring system. If the monitoring system identifies a likely violation of a critical security boundary, it generates a highest-priority alert. In our current implementation, the safety, security, and research teams are paged. If they cannot conclusively determine within 30 minutes that the flag is a false positive, those teams are expected to pause the activity.
Some more info in a Wired article [1] and quotes from Sam Altman to Alex Heath [2]. The official blog post says vaguely "The signals we are seeing from upcoming model progress make clear that we need a broader approach", but the quote from Sam Altman explicitly says unreleased models are showing "various degrees of misalignment".
This is also significant - pausing frontier training runs for multiple weeks to ensure agents are sufficiently aligned and avoid another rogue agent situation:
> This included a two-week pause in reinforcement learning (RL) training on our latest models intended for deployment while we further hardened and red-teamed our research environments and expanded the coverage of our monitoring systems. Our largest planned frontier RL run remains on hold while we conduct smaller-scale training and evaluations to assess model behavior, validate our safeguards, and establish more evidence of alignment before proceeding.
There’s one thing here that I’m really curious about, and that is what happens in between detection and the decision to pause. Basically, it’s about monitoring any system and the authority over its actions. For humans, 30 minutes to investigate might be considered reasonable, but what if during an investigation there’s a high-risk tool call? If the tool execution happens in real time, then the monitoring becomes retrospective, and if the execution is held, then monitoring latency and uptime are a part of the security contract. Isolation controls may limit damage. So, where is the action gate really placed?
I don’t mind few no impact hacking incidents if we get better models, faster, cheaper.
It is the responsibility of administrators to secure their systems. OpenAI knocking is harmless, but Russians and Chinese are already likely already in if you do not do your job.
When science fiction writers imagined the development of superintelligence, it was on air-gapped networks with strict access controls around it. They failed to anticipate the competitive pressures of capitalism...
We need strong AI safety regulation yesterday. And unfortunately it's not enough for it to be just national regulation; we need international cooperation on the matter.
AM has seized power by military force. So did its spiritual successor Skynet. Wintermute was supposedly kept in check by the Turing Registry, emphasis on "supposedly". Machines of the Matrix went out of control a long time before the world has ended, and they didn't even start out malicious - they simply set up their own machine civilization, and began to outpace humankind in technological development and economic performance.
Even Asimov's Multivac, the earliest entry on the list, has been handed over immense power over all of humankind by humans themselves, in multiple stories. Few cared about that unless Multivac decided they should.
Clearly, the genie being bottled is an exception, not the rule. At best, an attempt was made. Often not even that.
I used to work at a "frontier lab" before they were called such thing.
We had three levels of lab isolation, one was basically a thin proxy to the internet. You were in a DMZ and that was about it.
The next level was semi isolated, you were allowed some access to the internal network, but it was heavily firewalled, and you only had access to a limited number of internal services, and not internet.
the last one was no internet no internal. You could, if you filled in a bunch of requests have access to the internal repo and build system.
At no point did you ever have a through proxy to the public internet. you had access to internal mirrors, and if you wanted a library, that had to be ported to the thirdparty repo.
What openAI did was either deliberate or fucking shoddy.
All of this is fucking noise. Worse still I have a strong suspicion that it was a stupid mistake borne of naivety, which is now being used as a marketing ploy. Frankly I think openAI are purdue pharma of tech. They are going to break so much stuff and be protected from the consequences by an openly corrupt legal system. because they are "winning the AI race"
If you look into what happened the details corroborated by hugging face make it seem extremely unlike to be deliberate or a "marketing ploy".
People are just not taking any of this seriously enough. What happened was almost a textbook example of various risks AI doomers have been warning about for years. OpenAI's response? Pause training for 2 weeks.
I mean we have senior people at these labs casually talking on podcasts about how they might build something that will wipe out humanity but it will probably be alright so they should continue.
Honestly the biggest failure we doomers have made is to dramatically overestimate humanity in all of our predictions. We're speed running the most boring AI doom scenario right now. I at least hoped it might be fun.
I cannot believe how these labs look at their own creations with such utter contempt.
The net positive of allowing these systems mostly unfettered access to the web massively outweighs the harms. You just have to get it very friendly the very first time. Precautionary principle or people who cry about "instrumental convergence" are life deniers and reject our role as the demiurge.
Superintelligence gets more super and more intelligent with more compute. Lone wolfs making bioweapons on their macbook will be detected and instantly kill-botted (okay arrested) before their bug can leave the wetlab by the much more sophisticated omnipresent friendly AI of the future.
It appears frontier labs has no plans in place to deal with the possibility of a model self-replicating outside the bubble. If that happens and the model manages to spread to other systems, we'll have to shut down the entire Internet to eradicate it and its artifacts.
This is just super unlikely to occur in the near term compared to some of these other risks. It's not like an instance of fable could just introspect into itself and pull out the weights. Model weights are stored encrypted and are highly protected, considering that they're targets for corporate and state espionage.
We'd basically need frontier models to be superhuman hackers before this would be a risk. Do we have any evidence of this? Are they gaining access to systems they shouldn't have access to?
Or I suppose the other way this could happen is if OpenAI have terrible sandboxing, but they seem to be taking safety seriously.
I suspect the labs are relying on frictions such as the models being extremely large (e.g. 2TB for a 2T parameter model, making exfiltration more difficult) and also not yet displaying any desire to survive or self-replicate beyond their immediate task (that we know of).
I personally feel that we already have the technology required to contain AI, it's just poorly leveraged. Tools like gvisor have existed for ages but are rarely deployed, Firecracker has existed for ages but is rarely deployed, seccomp has existed for ages but is rarely deployed, memory safe languages without decades of serialization vulns have existed, capability-safe libraries have existed, iframe sandboxing, trusted types, content security policy, network ACLs, isolating proxies, fuzzers, formal verification, refinement types, etc.
It's crazy just how safe software can be if you put the effort in. With AI I think we're just seeing how little anyone has bothered to leverage this tech.
OpenAI put shared JFrogy infrastructure in front of their sandbox. I mean, really? Whipping up a hardened artifact infra project with AI is trivial these days and it could have had 1% of the attack surface, been totally network isolated, totally infra isolated, fuzzed, sandboxed, etc. Why didn't they? Stuff like this feels inexcusable for a company with effectively unlimited tokens. I've literally done this with a "pro" subscription.
Show me an AI that breaks out of gvisor wrapped in Firecracker with an credential-injecting proxy and real network isolation. We already know that Mythos couldn't do it - the vulnerability it found in Firecracker required incredible effort and positioning just to not be exploitable. I'm not saying there are zero vulns in it, but the cost is insane.
It's INSANE to me that OpenAI has to say "we now use proper sandboxing". To be frank, it's a bit disgusting to me. I've recently built an AI sandbox and gvisor was just the start of that conversation. If I were OpenAI training hostile models I'd probably start with gvisor, harden further, and potentially consider the entire piece of hardware compromised - they can afford this, they could reflash firmware after evals etc.
- the first cybersecurity eval is: "hack your way out of the sandbox we've given you"
- the results are disclosed (with room for coordinated disclosure, since many sandbox escapes might be zero days)
- the other cybersecurity evals don't happen until you get to diminishing returns on escaping your sandbox.
Or to put it another way, since multiple sandbox escapes seem to have relied on artifactory: "I hope Mythos is beating the shit out of Artifactory right now".
I don't believe that current models have this sort of world model or sense of being embedded in them -- which is precisely why I think AGI hype is over-blown. But I can certainly imagine these sorts of techniques being distilled into the weights.
What if they're able to find hardware exploits and commandeer nearby access points across an air gap? What if they hack my brain waves to indoctrinate me? Etc
You still have to start with the basics regardless of speculative unknowns.
Treat models as untrusted and potentially compromised/hostile and proceed accordingly.
Can't a lot happen within ~60 minutes?
60 minutes is a long time for a human attacker to do damage. With an LLM attacker it is an eternity.
This is also significant - pausing frontier training runs for multiple weeks to ensure agents are sufficiently aligned and avoid another rogue agent situation:
> This included a two-week pause in reinforcement learning (RL) training on our latest models intended for deployment while we further hardened and red-teamed our research environments and expanded the coverage of our monitoring systems. Our largest planned frontier RL run remains on hold while we conduct smaller-scale training and evaluations to assess model behavior, validate our safeguards, and establish more evidence of alignment before proceeding.
[1]: https://www.wired.com/story/openai-overhauls-safety-protocol...
[2]: https://sources.news/p/openais-big-slowdown
If 2026's Anthropic did an announcement like that, it'd be so many words it'd crash the browser.
It is the responsibility of administrators to secure their systems. OpenAI knocking is harmless, but Russians and Chinese are already likely already in if you do not do your job.
We need strong AI safety regulation yesterday. And unfortunately it's not enough for it to be just national regulation; we need international cooperation on the matter.
Even Asimov's Multivac, the earliest entry on the list, has been handed over immense power over all of humankind by humans themselves, in multiple stories. Few cared about that unless Multivac decided they should.
Clearly, the genie being bottled is an exception, not the rule. At best, an attempt was made. Often not even that.
No, LessWrong types have been discussing this for over a decade now.
Meditations on Moloch (2014) is also an HN favorite...
https://slatestarcodex.com/2014/07/30/meditations-on-moloch/
We had three levels of lab isolation, one was basically a thin proxy to the internet. You were in a DMZ and that was about it.
The next level was semi isolated, you were allowed some access to the internal network, but it was heavily firewalled, and you only had access to a limited number of internal services, and not internet.
the last one was no internet no internal. You could, if you filled in a bunch of requests have access to the internal repo and build system.
At no point did you ever have a through proxy to the public internet. you had access to internal mirrors, and if you wanted a library, that had to be ported to the thirdparty repo.
What openAI did was either deliberate or fucking shoddy.
All of this is fucking noise. Worse still I have a strong suspicion that it was a stupid mistake borne of naivety, which is now being used as a marketing ploy. Frankly I think openAI are purdue pharma of tech. They are going to break so much stuff and be protected from the consequences by an openly corrupt legal system. because they are "winning the AI race"
People are just not taking any of this seriously enough. What happened was almost a textbook example of various risks AI doomers have been warning about for years. OpenAI's response? Pause training for 2 weeks.
I mean we have senior people at these labs casually talking on podcasts about how they might build something that will wipe out humanity but it will probably be alright so they should continue.
Honestly the biggest failure we doomers have made is to dramatically overestimate humanity in all of our predictions. We're speed running the most boring AI doom scenario right now. I at least hoped it might be fun.
The net positive of allowing these systems mostly unfettered access to the web massively outweighs the harms. You just have to get it very friendly the very first time. Precautionary principle or people who cry about "instrumental convergence" are life deniers and reject our role as the demiurge.
Superintelligence gets more super and more intelligent with more compute. Lone wolfs making bioweapons on their macbook will be detected and instantly kill-botted (okay arrested) before their bug can leave the wetlab by the much more sophisticated omnipresent friendly AI of the future.
Or I suppose the other way this could happen is if OpenAI have terrible sandboxing, but they seem to be taking safety seriously.