Anthropic publishes its playbook for building AI safety evaluations
A 13 November research post sets out how Anthropic designs, sandboxes, grades and reports safety evals, from multi-turn simulated users to checking LLM graders against human experts.