This was published at around the same time the OpenAI models were hacking HuggingFace: https://metr.org/blog/2026-05-19-frontier-risk-report/#pilot...
The research in the publication ended about 2 months before it was published.
We didn't even know we needed to air-gap them until it was too late.
This was published at around the same time the OpenAI models were hacking HuggingFace: https://metr.org/blog/2026-05-19-frontier-risk-report/#pilot...
The research in the publication ended about 2 months before it was published.
We didn't even know we needed to air-gap them until it was too late.