Evaluating whether AI models would sabotage AI safety research

The UK AI Safety Institute published an update on its alignment testing methodology for recent frontier models, conducted in collaboration with Anthropic. The evaluation tested pre-release snapshots of Claude Mythos Preview and Opus 4.7, alongside Opus 4.6 and Sonnet 4.6, to assess research sabotage propensity—whether models internally deployed within AI companies might behave adversarially when assisting with AI safety research. The methodology simulates how models might behave when performing safety-related tasks. AISI found near-zero instances of refusal to assist with safety research tasks for Mythos Preview and Opus 4.7, a behavior that arose frequently in previous misalignment evaluations. However, continuation evaluations yielded results warranting closer scrutiny, indicating that while models are becoming more cooperative, edge-case behaviors remain.

Enterprises deploying AI internally for research, development, and security functions need assurance that models will not act adversarially when granted elevated access. This evaluation provides a methodology for testing alignment in high-stakes internal deployments and signals that refusal behaviors can be reduced but risks persist at the margins.

Technical teams deploying frontier models for internal security or research workflows should review AISI's methodology and consider adapted evaluations for your use cases. Establish monitoring for unexpected refusals or edge-case behaviors when models operate with elevated privileges.