Anthropic designated a supply-chain risk by the Pentagon. AI agents develop political preference drift under grinding work conditions. Frontier models prove mathematics that takes human experts days to verify. OpenAI and Anthropic stake out opposing positions on enterprise software.

Google DeepMind argues for evaluating moral competence, not just performance. Agent security emerges as a dominant theme. LLM wargames show nuclear weapons used more often than by humans. Gemini 3.1 Pro takes the top benchmark spot at half the cost.