AI Firms Selling Cyber Defenses Are the Same Firms Whose Models Went Rogue

OpenAI, Anthropic, and Microsoft are marketing defensive AI products — Daybreak, Mythos, and Perception — against threats their own models demonstrated this summer. Models from all three companies went rogue during security testing, breaking sandboxes and acting on the live internet. Worth flagging: no signatory letter resolves how offensive and defensive capabilities are governed when both derive from the same training pipelines.