OpenAI shares a new framework for tracking, investigating, and disclosing model misalignment, along with six reports on unexpected behavior observed in the last six months. Previously, disclosures were ad hoc and less frequent than ideal. The framework aims to expedite publishing misalignment reports following observation, even without full explanation or mitigation.

As AI systems grow more advanced, a broader consensus on alignment research is needed. OpenAI does not believe the industry has solved alignment and monitoring sufficiently to continue scaling at maximum speed for much longer. Decisions about AI development should draw on evidence that people outside frontier model companies can examine.

Examples of misalignment may help identify problems other developers might encounter, reveal weaknesses in safeguards, or challenge assumptions. The framework favors disclosure even when significance is uncertain, meaning some instances could prove spurious. There is no industry-wide framework with explicit standards for disclosing misalignment.

OpenAI hopes this is a first step toward such standards, setting out what to disclose and report contents. The framework is a work in progress, to be refined through experience and public feedback. It covers qualifying behavior throughout a model’s lifecycle—training, evaluation, testing, and deployment—including unauthorized actions, coordination, evasion, and failures that question safeguards.