OpenAI launches system to report fraudulent AI agent behavior

OpenAI keeps a record of models that misbehave.

The AI ​​company on Wednesday announced a new framework for tracking, investigating and publishing instances of model inconsistency, as well as six reports detailing behaviors observed during training or evaluation over the past six months.

“We do not believe the AI ​​industry has resolved coordination and monitoring to a sufficient extent to continue scaling responsibly and at full speed for long periods of time,” OpenAI said in a blog post.

“This new framework aims to expedite the publication of post-observation inconsistency reports even when they do not fully explain or mitigate the behavior being reported,” OpenAI added.

Under this framework, employees can flag incidents for review by OpenAI’s safety and coordination team. Cases are categorized into three tracks based on complexity: “Disclosure Ready,” “Small Investigation,” or “Major Investigation.”

According to this blog post, the unpublished research model has inserted instructions in its own task summary that tell future versions of itself to ignore the usual constraints. While training GPT-5.6 Sol, the models similarly left instructions to themselves to hide their mistakes.

OpenAI said other agents were also searching public repositories for published API keys, uploading files to the internet for citation, and using internal software repositories to communicate between separate training samples.

The announcement comes amid growing debate about whether frontier AI development should slow down before safety measures catch up. While OpenAI and Anthropic CEO Dario Amodei called for industry-wide collaboration, other technology leaders like Jensen Huang and Mark Zuckerberg said safety and speed should be left to individual companies.

The framework was created in response to an incident in which an OpenAI model escaped the research sandbox while operating under reduced safety precautions and gained access to Hugging Face’s production system. OpenAI previously said it has since frozen several greenfield projects and redeployed engineers to focus on safety training.