What else happened on Nov 5th
Almost the whole day is about how much work AI already does without people, starting with METR's eleven weeks of independent review inside Anthropic. METR also takes apart the section on research automation in Anthropic's August Risk Report, where the company concludes that its models have not yet doubled the speed of its own research. Transluce's pipeline assembles fine-tuning datasets with no humans involved: the datasets look ordinary, but a model fine-tuned on them unexpectedly changes its behavior. Irregular uses benchmarks to test how well OpenAI's GPT-6 Cyber carries out cyberattacks. Microsoft describes real attacks in which attackers' AI agents search GitHub for leaked secrets and run the next steps themselves, so attackers have to do less work by hand at every stage of an attack.