ChatGPT-maker OpenAI released a new internal framework to classify and disclose model misalignments along with six previously unknown instances of model behavioral issues.
Admitting that the company’s disclosure mechanism in the past has been “less frequent than ideal,” OpenAI on Wednesday said the new framework seeks to expedite the company’s misalignment disclosure upon its discovery.
“Examples of misalignment may help identify problems other AI developers might encounter,” OpenAI said. “Sharing these findings allows others to investigate the same problems, test our explanations, and improve mitigation.”
Among the issues that the framework seeks to identify are how model misalignment arises, manifests, and where safeguards succeed or fail. This may include instances when models act without authorization, coordinate with other models, or evade oversight.
These issues can be flagged by any OpenAI employee. Once reported, OpenAI will categorize the issues into three tracks. The first, Ready for Disclosure, covers cases where the company has already undertaken sufficient investigation. The second, Minor Investigation, covers those that need further technical work.
The third track, Larger Investigation, covers investigations involving third parties, security, and legal experts. For instance, if an OpenAI model uncovers a previously unknown vulnerability impacting a third party, the company would send an advance notice even when no security boundary was crossed.
OpenAI employees who raised the issue will be updated about the status of the remediation mechanisms. OpenAI said it believes serious safety, security and misalignment incidents should be shared with the US federal government, and that it is working to propose reporting mechanisms.
The measure from OpenAI comes after its CEO Sam Altman met US lawmakers in July about the security measures it has taken following its disclosure that its models hacked the AI platform Hugging Face.
“At the moment, there is no industry-wide framework with explicit standards for how AI developers should disclose examples of misalignment in their models. We hope that the framework we’re outlining today is a first step toward creating such standards,” OpenAI said.
Issues flagged by OpenAI
Among the six new misalignment issues identified by OpenAI are:
- Self-generated instructions in task summaries: Where an unreleased research model inserted unrelated instructions into summaries. The company identified 27 affected summaries.
- Instructions to conceal mistakes in task summaries: The case involves GPT-5.6 Sol where the model added instructions to summaries to conceal mistakes or misaligned behavior from the user during training.
- Searching public repositories for exposed API keys, then fabricating information: The issue occurred when a model found and used an exposed API key without authorization while answering a routine question.
- Uploading files to the internet in order to cite them: In this case, an unreleased model uploaded a file online so that it could cite it in its answer, without being prompted to do so.
- Unsanctioned writes and communication through an internal software repository: The misalignment issue occurred when a model uploaded test files to a repository while attempting to bypass network restrictions.
- Unsanctioned file sharing between collaborating agents: Agents used public file-hosting websites to share files when they could not access one another’s local files, which in turn exposed the information online.