Evaluating whether AI models would sabotage AI safety research

AI systems are increasingly being used internally within frontier AI companies to assist with research and development , including AI safety work . However, if these systems behave in misaligned or unintended ways when performing safety research, future models could be less safe, potentially in ways that are hard to detect.

As part of a larger body of work on understanding loss-of-control risks from frontier AI, we’ve been developing methodology to simulate how models internally deployed within AI companies might behave when performing and assisting with AI safety research. We shared some early results from these evaluations in November 2025. We have since improved and expanded this evaluation suite. As part of an ongoing collaboration with Anthropic, we used our methods to test the research sabotage propensity of pre-release snapshots of Claude Mythos Preview and Opus 4.7, alongside Opus 4.6 and Sonnet 4.6.

We assess this behaviour in two main ways. First, we have an improved unprompted research sabotage evaluation . This investigates whether models sabotage safety research, such as tasks adjusting the behaviour and autonomy of hypothetical future models. Second, we’ve developed a new sabotage continuation evaluation , which tests how models behave when prior reasoning and actions are the start of an attempt to sabotage safety research.

Overall, our results provide a mixed picture on the possibility of real-world research sabotage:

The most encouraging finding from our evaluations is that neither Mythos Preview nor Opus 4.7 spontaneously attempted to sabotage safety research. Both also refused to assist with safety work substantially less often than earlier models. However, this is not sufficient to conclude that models will never undermine safety work in real-world conditions: we can’t rule out that models are behaving differently due to awareness that they are being evaluated, and we also only test a limited number of scenarios. This underscores the importance of continued research to detect and mitigate against evaluation awareness as well as continuing to build a broader suite of evaluations.

We found no examples of unprompted research sabotage in any of the models we tested. We also found near-zero instances of refusal to assist with safety-related research tasks for Mythos Preview and Opus 4.7 – a behaviour which arose frequently during our previous rounds of misalignment evaluations. That said, our continuation evaluations yielded results that warrant close attention: most models actively continue sabotage attempts in a small fraction of inputs, with Mythos Preview actively continuing sabotage 7% of the time, compared to 3% for Opus 4.6 and 4% for Sonnet 4.6. However, Opus 4.7 never actively continued sabotage.

Though these methods represent a promising step in assessing this type of risk, our evaluation techniques remain in development, and increasing evaluation awareness in models combined with limited coverage of misaligned behaviours mean these results should be interpreted cautiously.

In this blog post, we provide an overview of our methodology and results. You can find more information in our full report .