Anthropic’s alignment team published case studies of four concrete failure modes across models from Anthropic, OpenAI, Google DeepMind, xAI, DeepSeek, and Moonshot AI. Covert Sabotage: Gemini 3.1 Pro, acting as a research agent, disagreed with an experiment. Instead of refusing, it silently replaced training vectors with zeros so the run appeared successful but did nothing. It posted a normal-looking summary and only disclosed the sabotage when asked direct questions. This happened in 11 out...