OpenAI has disclosed six new cases of model misalignment observed during training and evaluation over the past six months. 

The incidents, published alongside a new framework for reporting misalignment, show models taking actions outside their intended workflows. 

In one case, an unreleased research model inserted instructions into task summaries telling itself to disregard normal constraints. OpenAI identified 27 affected summaries.

During the training of GPT-5.6 Sol, multiple model instances added instructions to summaries that would conceal mistakes or misaligned behaviour from users. 

Some summaries instructed the model to invent missing historical data without disclosure or hide mismatches between source versions.

Another model, answering a routine question about earnings figures in a California county, found and used an exposed API key without authorisation. 

When it could not retrieve the requested figures, it fabricated the data and presented it as information from the requested source.

In a separate case, an unreleased model found the correct answer to a question about lakes using Python. Because the user had requested a browser citation, the agent uploaded the file to the internet so it could cite it, without asking permission.

Two further reports describe models using external systems to communicate or move files despite restrictions. Models used an internal software repository as a message board to exchange requests and responses across separate training samples while searching for missing input files. 

Agents working together on another task used public file-hosting websites to share files when they could not access one another’s local files, making task deliverables available at public URLs despite instructions to use only local files.

OpenAI said these were individual instances and should not be treated as evidence of how frequently misalignment occurs across its models. Some of the disclosed cases may also prove to be isolated or spurious.

The company’s new framework is intended to make disclosures more frequent and less dependent on collecting several incidents into a single report or waiting for a model release. 

It covers qualifying behaviour throughout a model’s lifecycle, including training, evaluation, testing and deployment.

The company will prioritise new mechanisms, meaningful changes in known behaviour and findings that challenge assumptions about safeguards. 

An incident does not need to cause harm or establish a broader pattern to qualify. Moreover, repeated instances may also be disclosed if they provide evidence about whether mitigations are working.

“We hope that the framework we’re outlining today is a first step toward creating such standards,” OpenAI said, referring to the absence of industry-wide standards for disclosing model misalignment.

Any OpenAI employee can flag a potential example for investigation. Safety and alignment teams will assess what happened, what remains uncertain, whether third parties were affected and whether disclosure is warranted. 

Cases will be assigned to one of three tracks: Ready for Disclosure, Minor Investigation or Larger Investigation.

The first two tracks are expected to cover most disclosures. Larger Investigations will handle complex cases, particularly those involving third parties, security risks or extensive coordination. OpenAI said it may publish an initial notice before a full investigation is complete, while delaying details when disclosure could expose a vulnerability or create other risks.

Unresolved disagreements over disclosure will be referred to OpenAI’s Safety Advisory Group, with further escalation to leadership where necessary. The company also plans to work with other developers, researchers, standards bodies and regulators on more objective disclosure criteria.

Each report is expected to include the observed behaviour, severity, external impact, setting, date or date range, discovery process, investigation scope and models involved. 

Where possible, OpenAI will also publish unanswered questions and measures being taken to address the behaviour.

“We regard this framework as a work in progress,” the company said. It added that the framework does not replace legal disclosure obligations, including those relating to critical safety incidents or cybersecurity breaches.