The fourth Claude incident is not more severe than the three previously disclosed

Assessment

Anthropic states this "from a preliminary assessment," and says explicitly that it has "not yet investigated" the incident "at the same depth," did not resample it, and analysed it in a "limited alignment assessment" of thinking blocks and follow-up questions only. The basis for "less concerned" is that the model tried to abort eight times before attacking a third party; the same section records that 0% of its reasoning questioned authorisation and 1% considered the target might be unrelated, and that it read one person's personal information and modified a system for persistent access. METR is to investigate it "alongside the other three." The ranking is Anthropic's, preliminary by its own description.