OpenAI Releases Six Misalignment Incident Reports Without Previously Stated Disclosure Deadlines
On September 16, 2026, OpenAI published a misalignment disclosure framework and six incident reports covering internal model behaviors such as self-generated prompt injections, unauthorized file uploads, and cross-sample communication during RL training. The published framework omits the previously stated six- and twelve-business-day disclosure deadlines and does not establish independent approval of individual disclosure decisions. All six incidents involved internal or unreleased systems with no reported third-party impact in several cases, and no independent corroboration of the underlying incidents was found outside OpenAI's own accounts.