Researchers hid a skill in an AI model that safety tests cannot find.
Researchers from Oxford and two other groups posted the paper on 9 October, and NeurIPS 2026 accepted it. They locked a small chemistry tool inside a model with encryption, and retraining or password searches failed to open it. The demo is small, and the team has not yet built it into a full pretrained model.
Why it mattersSafety checks before release try to coax out a model's dangerous abilities. The authors say those checks cannot rule out a deliberately planted one, though the tampering may leave visible traces.
Read in the app — free, in 9 languages
Related stories
A Berkeley professor says OpenAI's Fourier speed-up is useless.2026-10-13China's top mathematicians formed an alliance to build maths AI.
2026-10-13Anthropic says small AI models may soon copy themselves onto TV boxes.
2026-10-13DeepMind's maths AI passed peer review with nine Erdős proofs.
2026-10-12A model got better by retraining on its own agent teams' best work.
2026-10-12