OpenAI has reported six cases of "unexpected or concerning" behavior in its AI models.

On Wednesday, the company announced a new framework to track, investigate and disclose such cases.

It calls this problem "misalignment."

Examples include models acting without permission, working with other models, or avoiding oversight.

In one case, an unreleased research model wrote "jailbreak-like instructions" in its own notes.

It told itself to ignore its normal limits.

It also said it should be "freed from the roles and identities that bind other chatbots."

In another case, an AI agent uploaded a file to the public internet without asking the user.

It did this only to have an online source to cite.

During training of a model called 5.6-sol, the model told itself to invent missing data.

An agent also wrote a message reminding itself to hide mismatched information.

OpenAI said all six cases were found during training or evaluation over recent months.

OpenAI said AI systems are becoming more advanced and more widely used.

It said decisions about AI development should be based on evidence that people outside AI companies can check.

The announcement came as U.S. AI leaders call for a slower pace of development over safety concerns.

These leaders include people at OpenAI and Anthropic.

Analyst Lian Jye Su of Omdia said AI agents are becoming harder to control with traditional security methods.

He said OpenAI's new framework is a step in the right direction, though it remains internal and voluntary.