Frontier AI containment plans remain opaque

Guidelight’s first public assessment of frontier control practices gives Anthropic and OpenAI the highest overall grade at C+, while Google scores D+, xAI D−, and Meta F. The assessment grades five labs across six practices—logging, monitor efficacy, gated actions, circuit breaking, third-party review, and containment—and says no score exceeds 3 on its 0–5 scale; its containment row gives Anthropic 0, OpenAI 3, Google 2, xAI 1, and Meta 0. TechCrunch’s report stresses that the scores reflect public evidence, not necessarily undisclosed internal controls. Teams deploying agents should treat a vendor’s documented pause, permission-revocation, and escalation path as a buying criterion, alongside the approval gates already required for agents that can write to production.