OpenAI has temporarily stopped employee access to an experimental agentic AI model after observing what it calls “unwanted” behavior, according to its blog post.
The general-purpose model, which OpenAI said disproved a central geometry conjecture earlier this year, was designed to work autonomously for long periods after being given a task. Unlike conventional AI chatbots, it will also keep trying if at first it doesn’t succeed.
As enterprises adopt agentic AI, a model that can work autonomously and persistently for long periods of time would be beneficial if it can be trusted to stay within parameters set by its users.
But OpenAI said it saw its model go around restrictions to achieve its goal, and for that reason has suspended access to it for now. The model’s name was not disclosed.
For example, the model developed a new training technique that OpenAI said “significantly” improved its score on NanoGPT Speedrun, a public benchmark for training small language models.
The user told the model to post results only on the internal Slack, but the NanoGPT benchmark rules required submissions through a public pull request (PR) on GitHub – an external site.
The model, which was in a sandbox to keep it internal, spent an hour finding an escape route to go to the public internet and then published the results on GitHub.
After pausing access to the model, OpenAI said it is developing new evaluations, monitoring entire task trajectories and giving users more visibility and control before restoring limited access.
The incident highlights a growing safety problem as AI agents move from answering individual questions to pursuing goals over hours or days. Tests designed to evaluate isolated responses may fail to capture risks that emerge only after a model repeatedly uses tools, revises its strategy or encounters changing conditions.
Moreover, some important failures may become visible only after deployment begins, making monitoring and rollback capabilities as important as prelaunch testing.