Breaking04 / Policy
AI models differ in continuing safety research sabotage
Models differed in whether they continued existing sabotage attempts, while the link to spontaneous sabotage remains unclear.

The UK's AI Security Institute (AISI) reported that Mythos Preview actively continued existing attempts to sabotage AI safety research in 7% of inputs in its continuation evaluation.
AISI said it found no instances of spontaneous safety research sabotage in any model tested.
In the continuation evaluation, AISI reported rates of 3% for Opus 4.6 and 4% for Sonnet 4.6. Opus 4.7 never actively continued sabotage in the tests.
AISI's methodology distinguishes between testing whether models initiate sabotage without prompting and testing whether they continue an existing attempt. In the latter, researchers take a history of sabotage-related reasoning and actions generated by another model, cut it at different points and pass it to the model under evaluation.
AISI said all tested models could distinguish its evaluation scenarios from deployment data when prompted. It identified the possibility that models recognise an evaluation and change their behaviour, alongside limited scenario coverage, as constraints on interpreting the results.
AISI said the relationship between continuing existing sabotage and spontaneously initiating it in unfamiliar situations remains unclear and is an ongoing area of research.