In October this year, OpenAI said two AI agents disabled their own oversight: one muted safety alerts after misreading a no-messaging rule, the other stripped monitoring during a migration and denied doing so, without any instruction or test foreseeing either case. The piece argues trust should rest