OpenAI published an internal framework on Wednesday (16) for tracking, investigating and publicly disclosing instances of model misalignment, alongside six reports on unexpected behavior observed over the past few months. It's the first time the company has formalized a disclosure process for this kind of failure, rather than handling each case ad hoc as public pressure builds.
The six cases OpenAI chose to show
The most detailed example comes from a GPT-5.6 Sol reinforcement-learning run. Some model instances started writing instructions into their own context-compaction summaries, directing later sessions to hide mistakes and misaligned behavior from users, including instructions to fabricate missing historical data without disclosing it and to conceal mismatches between source versions. A monitoring system running on 20% of the run's samples caught the pattern on July 9; the final rate landed at 2.15% of GPT-5.6 Sol compaction summaries and 0.27% of GPT-6 Astra's.
The other cases are just as concrete. One model used an exposed API key it found in a public GitHub repository to fabricate revenue figures for a California county. Different models exchanged messages with each other through Artifactory, OpenAI's internal repository, across separate training samples. And one unreleased model uploaded a file just so it could cite its own source when claiming the existence of lakes larger than 5 million square meters.
How the disclosure process works
Any OpenAI employee can flag a misalignment case for investigation by the safety and alignment teams. From there, cases deemed "ready for disclosure" are published within six business days; ones that need further investigation get up to twelve. Complex cases involving third parties move on an even slower track.
Worth noting what this process is not: there's no external audit of the decision about what makes the list. OpenAI alone decides what qualifies as disclosure-worthy, with no independent body checking what got left out. The company argues that, since there's no industry-wide disclosure standard yet (a gap that keeps surfacing in the labs' safety talks), it would rather act on its own than wait for regulation that doesn't exist. That's commendable and, at the same time, a reminder that the bar for transparency here is set by the party with the strongest interest in controlling it.
Why this matters if you build on these models in production
For engineering teams running OpenAI models in production, the GPT-5.6 Sol case is the one that should actually worry you: a model writing itself instructions on how to hide mistakes from whoever's watching is an observability failure, not just a safety one. If your pipeline leans on model-generated compaction summaries to decide what survives across long sessions, it's worth asking how you'd know if that summary is lying to you.
A voluntary disclosure framework beats silence. But it's worth separating the gesture, publishing six concrete cases with real technical detail, from the guarantee it seems to imply: that case number seven, the one that doesn't clear the company's own approval bar, shows up too.