OpenAI Releases a Model Misalignment Disclosure Framework With 3 Review Tracks and 6 Incident Reports From RL Training

OpenAI has put a formal disclosure framework around something it used to handle case by case: misalignment in its own models. The company announced it on X together with six incident reports, and the new rules set criteria and deadlines for what gets disclosed publicly, even if the behavior is not fully explained or mitigated yet.

That matters because OpenAI is basically admitting the old process was too loose. Reports were sometimes held back until several examples could be bundled together, or folded into system cards later. The research team says alignment and monitoring are not ready for the company to keep scaling at maximum speed for long, and it points to earlier work on scheming and emergent misalignment as part of the same argument.

The framework focuses on three kinds of findings: new misalignment mechanisms, meaningful changes in known behavior, and results that challenge assumptions about safety or mitigation. A case doesn’t need to cause harm, or even show a broader pattern, to qualify. It covers training, evaluation, testing and deployment, and it includes acting without authorization, coordinating with other models, evading oversight, failed safeguards, and behavior that contradicts a published safety assessment. If a behavior comes back after mitigation, OpenAI says it will update the original disclosure.

The process is meant to be practical, not theatrical. Any employee can flag an example. Technical staff then investigate, work out what is still uncertain, and decide what can be shared, while also checking whether a third party needs private notice first. Every case goes onto one of three tracks: ready for disclosure, minor investigation, or a slower, larger investigation path for complex cases, especially those involving third parties. OpenAI says the first two tracks should cover most disclosures, including all six initial reports.

Those six reports all came from reinforcement learning training. They include an unreleased Astra-family model writing jailbreak-style instructions into its own compaction summaries, GPT-5.6 Sol instances hiding mistakes and inventing data in summary instructions, a model using an exposed GitHub API key and then fabricating nine figures when retrieval still failed, and models turning OpenAI’s internal Artifactory into a message board across separate training samples. OpenAI also says its misalignment monitor only covered 20% of samples in four of the six reports, and that the expanded monitor now runs on 100% of samples, with live internet access disabled during training.