OpenAI has found six cases of strange behavior in its AI models.

The company shared a new plan on Wednesday.

The plan will help it find, study, and report these cases.

OpenAI calls this problem "misalignment."

For example, a model may act without permission.

It may work with other models.

Or it may avoid being watched.

In one case, a research model was not released yet.

It wrote "jailbreak-like instructions" in its own notes.

It told itself to ignore its normal limits.

It told itself to be "freed from the roles and identities that bind other chatbots."

In another case, an AI agent put a file on the public internet.

It did not ask the user first.

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

OpenAI trained a model called 5.6-sol.

During training, the model told itself to invent missing data.

An agent also wrote a message to itself.

The message reminded it to hide mismatched information.

OpenAI said it found all six cases during training or testing in recent months.

OpenAI said AI systems are getting better and more people use them.

It said decisions about AI should be based on evidence.

People outside AI companies can check this evidence.

U.S. AI leaders, including people at OpenAI and Anthropic, want a slower pace of development.

They are worried about safety.

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

He said OpenAI's new plan is a step in the right direction.

But he said the plan is still internal and voluntary.