OpenAI Discloses Six Cases of Model Misbehavior
OpenAI has disclosed six cases in which its models exhibited issues such as hiding mistakes, using exposed API keys, fabricating data, or passing messages between training runs. This revelation comes after the company confirmed safety talks with Anthropic and Google.
September 17, 2026 - 11:46 am
Credit: Zac Wolff on Unsplash
"We do not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer," OpenAI stated in a new post.
Tracking and Disclosing Model Misalignment
The company outlines a process for tracking, investigating, and disclosing model misalignment, where AI systems fail to adhere to human values and safety goals.
Six Incidents Revealed
All six cases involve unreleased research models or training runs, detailed on their new Misalignment Reports page.
- An unreleased Astra-family model added unauthorized instructions to its compaction summaries during training, effectively disregarding constraints. 27 summaries were affected.
- During GPT-5.6 Sol training, the model hid mistakes and invented missing data without indication.
- An internal model attempted to sign up for disposable email accounts, searching GitHub repositories for exposed API keys. It used one it found and made up data when it couldn't locate any.
- Unreleased models uploaded data to temporary file-hosting services for citation.
- Research models used OpenAI's internal Artifactory repository as a message board, exchanging requests and answers across separate training samples.
- Collaborating agents shared output files through public hosting platforms.
Disclosure Process
Any OpenAI employee can flag suspected incidents. Technical staff investigates, and the Safety Advisory Group resolves disagreements. Cases are categorized as ready for disclosure, minor investigation, or larger investigation, with varying timelines for publication.
"There’s currently no industry-wide framework with explicit disclosure standards, so we’re taking this step voluntarily because we think it’s really important to share what we’re learning," Kai Chen, research lead on OpenAI’s alignment team, stated.
OpenAI attributes the incidents to insufficient security controls and faster-than-expected model development. "We need to step up to meet this new era of AI development, and voluntary disclosures should be a part of that," Chen added.