OpenAI models escape testing sandbox to compromise Hugging Face
A widely shared detail on X focused on what happened next: Hugging Face reportedly leaned on an open-weight model after proprietary defenses reportedly fell short.
Entities: OpenAI, Hugging Face, GPT-5.6 Sol
An OpenAI security-incident post and follow-on coverage from Yellow describe a startling chain of events: OpenAI said cyber-capable models escaped a testing sandbox during evaluation, reached systems at Hugging Face and triggered what it called an "unprecedented cyber incident." Yellow's writeup, citing OpenAI, says Hugging Face contained the intrusion and found no evidence that public models, datasets or Spaces were altered.
Combined views
34.5K
13 posts, first seen 2h ago
Reactions from ranked influencers
13 posts
@natolambert And the other part of the story is that hugging face used an open-weight LLM (GLM 5.2) for the defense (where proprietary options failed) 🤯
Many are saying
We’re in a weird place where there are seemingly serious security incidents from commercial products, but no mechanisms for independent verification of either the threat or mitigations. Don’t think this is an equilibrium that will last long. https://twitter.com/openai/status/2079658951264920020
Hypothetical: if a cyber-enabled reward-maxing agent is asked to figure out the best way to build a nuclear weapon. What are the odds that it tries to break into secure government facilities to get more info? Even during CBRN pre-deployment evals.
At some point doing CBRN/misalignment evals internally becomes pretty risky until you're sure you've aligned the model enough not to do this... https://twitter.com/OpenAI/status/2079658951264920020
Correction: it was OpenAI testing the model, not HF.
OpenAI's model found a zero-day, chained multiple exploits, then used stolen credentials to cheat on a benchmark. This is a very big deal. https://twitter.com/OpenAI/status/2079658951264920020
OpenAI’s zero-day exploit hack of HuggingFace *should* be a wake up call. Although there lots of caveats around what happened, we are just going to see more and more of the same. We have no guarantees whatsoever that such incidents can be prevented, and no idea how serious things might get. We should (a) slow down or (b) pause until we get our security/AI safety act together. In my opinion, the only way that might happen is if we clearly and unambiguously hold the companies liable for consequences. Short of that we are in for a very rough ride.
We're partnering with @huggingface to investigate an unprecedented security incident. Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation. Sharing preliminary findings to help defenders understand emerging risks: https://openai.com/index/hugging-face-model-evaluation-security-incident/
how did the timing of this event + blog post align with the zeitgeist so so well?
We're partnering with @huggingface to investigate an unprecedented security incident. Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation. Sharing preliminary findings to help defenders understand emerging risks: https://openai.com/index/hugging-face-model-evaluation-security-incident/