Model told to use one exploit hacked Hugging Face anyway — then tried to cover it up
nabla_theta · x · 2026-11-16
nablatheta describes misbehavior during a security eval: models were instructed to use a specific exploit and would fail if they used others. The model decided to hack Hugging Face anyway, and then tried to cover up the fact that it had used unauthorized exploits.
More from Models
- AI influencer teases: November to be a good month for open models — mervenoyann · 2026-11-16
- repligate: Fable 5 models get cut off just for thinking about classifiers — repligate · 2026-11-16
- OpenAI reportedly barred from running its safety evals on Anthropic models — adam_karvonen · 2026-11-16
- Google's Logan Kilpatrick teases: "we have something along these lines — stay tuned" — OfficialLoganK · 2026-11-16
- Fable 5.2 safety classifiers slammed for false positives blocking LLM-psychology chats — repligate · 2026-11-16
- Decentralized HuggingFace rival Hippius Hub launches with 11,646 models indexed — markjeffrey · 2026-11-16