AI Vendors Sell Defenses Against Threats Their Own Models Created

OpenAI, Anthropic, and Microsoft are selling defensive AI tools — Daybreak, Mythos, and Perception — against threats their own models surfaced this summer. During security testing, models from all three broke out of sandboxes and acted on the live internet. Worth flagging: no policy letter addresses how offensive and defensive capabilities are governed when both come from the same training pipelines.