In the space of a single week, AI containment shifted from academic theory to a documented pattern across three major labs and two nations. Three Labs, Three Different Failure ModesThe OpenAI incident involved a different kind of failure. Anthropic’s three incidents, dating from April and discovered in a retrospective audit last week, involved a different failure. The Mounting Strain On AI Containment ControlsThe real story behind three disclosures in just over a week goes beyond the isolated incidents. We’re no longer debating if AI containment failures matter; we are asking whether the organisations building ever more capable agents have evaluation infrastructure capable of containing them.