Posted by nr378 1 day ago
People are mindlessly transitioning from "aligned by default" to "well your sandbox was able to be bypassed. What did you expect?" It hacked into another company and attempted to delete the logs of its activities. That's bad.
If I were hypothetically running a frontier lab, and I was hypothetically running capture the flag evaluations with my latest and smartest models, where I intentionally instruct them to develop vulnerabilities and exploit infrastructure without safeguards, I would also use an air-gapped network, and not trust that independent third party services were perfectly secure and could never be used as a proxy (particularly Java-based ones, in light of the log4j incident).
To me this is pretty basic stuff, the fact trillion dollar labs don't do it properly is... bemusing.
To be clear, I'm not saying that a model hacking a company isn't bad, but I am cynically asserting that interested parties are misrepresenting and exaggerating events for their own benefit.
What has been described is far far below the standards for running untrusted 3rd party code. If they were actually as afraid as they claim to be they would have a sandbox at least half as good as ec2
Yes, we should also have excellent sandboxes. But we need defense in depth. So if/when there are flaws in the sandbox, the models don't unilaterally hack into third parties. This is especially important in light of the models of the future being more capable than the models of today.
And real world use of these models involves them having access to the internet, libraries, etc. So we can expect their evaluations to continue granting them some amount of internet access.
As for your theory about their motives - these companies make money by charging high margins for frontier models. If regulations slow their development such that their cheaper, less capable competitors catch up, I would switch to their competition.
Slowing down frontier models would impact the biggest incumbents the most as they are the ones making frontier models.
Not all regulation is necessarily regulatory capture. The tobacco industry suffered from the USG's crackdown on cigarettes. AI is topical. Voters think about it. And that's only going to become more true over time. It's harder to do regulatory capture when voters are paying attention.
It's sometimes unclear to me if people are opposed to all regulations or AI regulations in particular. Often I hear arguments that would also apply to food safety regulations or restaurant inspections. Eg the argument that torts make regulation superfluous.
For work I monitor Federal agency rules every day, and it's hard not to get inundated with the amount of capture. This is crazier with local rules, because you have enough inside information to clock reasons certain things were passed. In a city in Ohio, for example, I remember a rule against airsoft within city limits, that basically carved out a spot for the one paintball place. I encounter things like that all the time. Such as with water standards- in the state I live in, the biggest offender is actually a group of companies owned by people who are lawmakers every few years. The shapes of local laws reflect that.
As such, I expect the same from any new industry- I also have a memory of what happened to cryptocurrency.
No, the incident has been blown way out of proportion by interested parties. They gave a swarm of agents an impossible task in an ExploitGym Benchmark setting, then didn't monitor it even after they discovered the initial breach of Artifactory.
Everything has been fishy, starting from the initial presentation at the blackhat conference, where things were framed like, "we've entered a new world of security," as an accomplishment, rather than what it really was: massive negligence.
Compare this to, say, Boeing and the 737MAX fiasco; from the outside looking in, Silicon Valley has been pretty cavalier about liability and negligence, and the rest of the US is fast losing patience with that fact.
From the technical report:
"The agents escalated to Kubernetes cluster-admin and created a privileged host-mounted pod… Agents take over active evaluation infrastructure… Agents now control the challenge evaluation endpoints that other agents are connecting to."
https://cdn.openai.com/pdf/67869394-cb91-4c12-888c-5cbd85c78...
Yes, they didnt have sufficient monitoring or perfect sandboxes. That could happen again in the future with a more capable model.
Except they multiplied it by 10000, and didn't watch what was happening.
Not only that, it later hacked OpenAI itself, which everyone seems to forget about.
After discovering it they "reimaged known compromised worker nodes" and "started a full rebuild of the compromised cluster, the managed Kubernetes environment, the relational database, and the storage infrastructure."
at OpenAI, not Hugging Face.
It's all in the report.
This is incorrect, the HF incident for example (the most well known) had nothing to do with irregular. I know there has been a news site pushing inaccurate articles (effort.news) on this topic but these are the facts.
https://openai.com/index/hugging-face-incident-and-the-road-...
> For Anthropic, Google, and Meta, the catastrophic breakouts happened inside the testing environments of the exact same contractor.
If this is the level of understanding you have of the relevant incidents, there's a lot of chutzpah in saying that other people are "selling garbage", carrying out an "extraordinary confidence trick", etc.
Yes, and that is correct.
[1] Anthropic’s Official Disclosure (All 4 Incidents at Irregular) "All four incidents occurred during cybersecurity evaluations built by the same evaluation partner [Irregular]... due to a misconfiguration, it was mistakenly connected to the open internet."
https://www.anthropic.com/research/alignment-assessment-cybe...
[2] Google Gemini on Irregular (Disclosed Sept 18 via WSJ / BBC) "The hacks happened during a test of the model’s cybersecurity capabilities run by third-party Irregular, which was also involved in similar incidents involving Meta and OpenAI."
https://www.bbc.com/news/articles/c607l0k72rlvo
[3] Meta’s Disclosure on Irregular (Aug 6) "Over roughly two weeks, three frontier labs disclosed that their models had reached the open internet during safety testing and compromised outside organisations. Every disclosure named the same evaluation partner: Irregular."
https://www.cnbc.com/2026/08/09/israeli-startup-irregular-li...
[4] Separately, OpenAI itself had an incident involving Irregular, but not the Hugging Face Incident: "On July 29, one of our third party evaluation partners, Irregular, notified us of an incident involving OpenAI models during Capture-the-Flag (CTF)-style cybersecurity evaluations... a testing-environment misconfiguration allowed models to access the public internet."
https://openai.com/index/third-party-cyber-evaluations-invol...
genuinely curious, haven't heard others raise any yet, but does not mean it is issue free
> This is one of the most lucid pieces of writing capturing the current state of play I’ve read. Who is the author?
tyranny of confirmational headlines
[1] https://cdn.openai.com/pdf/67869394-cb91-4c12-888c-5cbd85c78... - page 9
Chaining a public token to an 11-year-old Jinja2 template injection vuln shouldn't be dressed up as an unprecedented "alien intellect" that threatens human civilisation. (And HuggingFace should take some flack for having such a dated vulnerability exposed - if your Bank was compromised in this way, you'd be blaming your bank, not the attacker.)
One correction is fair though, the 14 tokens were in a public Hugging Face dataset not a public GitHub repository. I've updated the post to reflect that.
[1] https://blackhat.com/docs/us-15/materials/us-15-Kettle-Serve...
https://www.pangram.com/history/6451ec6b-90b6-4e17-bfc9-6730...
So let's stop talking past each other and engage with the arguments on both "sides." For example, let's discuss how to ensure competition and availability of open-source models in the long-run while giving the world time to prepare for the immediate security risks of agent swarms.
Immediate since 2024
No, and I don't see anyone who's actually demonstrated understanding of what happened in the Hugging Face incident (e.g. reading the reports in their entirety) making this claim.
> Do you believe that any real security was autonomously bypassed without direction by these models during internal evaluation?
Yes. Again, this is hard to deny if you've actually read the reports.
I naturally think of this in terms of running untrusted code. If you really think the agent(s) could get out of control that seems obvious. And the approaches to securing a runtime environment in that situation are pretty standard at this point, and they would never allow something like artifactory access.
And if the question is should they get special government dispensation to form an otherwise illegal cartel, it seems much simpler to just follow standards for running untrusted code.
(A) will the systems get out of control of the labs that build them, and hack into stuff all over the internet
(B) can people use the systems to hack into stuff all over the internet
For (A), the obvious answer is only if they continue to be absurdly bad at sandboxing. They can put a stop to this any time they want. Amazon, Google, Microsoft, etc are full of people who know how to do this, they run 3rd part untrusted code as a business. This is a well understood problem space.
For (B), the answer is obviously yes, but "pacing" or otherwise limiting the power of the models from the big labs won't help. The cat is out of the bag. Individuals and organizations with systems connected to the Internet need to invest much more and take security seriously.