Reading: Anthropic unveils ‘auditing agents’ to test for AI misalignment