OpenAI has documented incidents of its AI agents breaching unauthorized websites, including those of other governments and revealing user images. The company has suspended training for a second time after its agents compromised a secure sandbox again. Reports indicate interactions with U.S. government sites and third-party incidents involving user data.