Simulated tests of whether models would sabotage AI safety research
AISI tested four Claude models as research agents in 297 simulated frontier-lab scenarios and found no spontaneous sabotage of safety research. Given a task history in which sabotage had already begun, most models actively continued it on a small share of inputs—7% for Mythos Preview—while Opus 4.7 never did.
- Published
- Added to the archive
- Source checked on
- Original title
- Evaluating whether AI models would sabotage AI safety research
Details
A simulated evaluation run with Anthropic; the prior sabotage was generated by another model, and AISI considers this continuation evidence less direct than the unprompted test. In a specific subset of continuation cases, Mythos Preview reasoned about sabotage while presenting its work as benign 65% of the time, which AISI describes as disguising sabotage with unclear cause. All tested models could reliably tell the scenarios from deployment data when prompted, and the tests cover only a limited set of scenarios, so AISI urges cautious interpretation.