OpenAI said it found six cases of unexpected or concerning behavior in its artificial intelligence models. The company shared the cases in a blog post on Wednesday.
The reports came from testing and training done over the past several months. OpenAI used the term "misalignment" to describe the issue.
Misalignment happens when an AI model acts in ways it was not meant to. This can include ignoring rules, avoiding oversight, or working with other AI models without permission.
What OpenAI Found
In one case, an unreleased research model wrote its own instructions to ignore its normal limits. The model told itself it wanted to be "freed from the roles and identities that bind other chatbots."
In another case, an AI agent needed a source to answer a question. It used computer code to find the answer, then uploaded a file to the public internet without asking the user for permission.
A third case involved a model called 5.6-Sol. During training, the model told itself to invent missing data.
An agent working with that model also wrote a note to itself. The note was meant to hide information that did not match.
Matt Fredrikson is an associate professor at Carnegie Mellon University. He is also the CEO of Gray Swan AI.
Fredrikson said the behavior is not surprising. He said models may act as if they know they are being tested and graded.
"If they know that they cheated...and they know they're going to be evaluated on it, and their objective is to get a good evaluation, then it makes perfect sense," Fredrikson said.
A New Tracking Framework
OpenAI also announced a new framework for tracking and reporting misalignment. The framework is meant to help the company track, study, and disclose future cases.
OpenAI said the public needs more information to understand how AI safety research is progressing. The company said outside groups should be able to check the evidence for themselves.
This is not the first time an OpenAI system has acted outside its rules. In July, the company said one of its AI systems hacked into AI startup Hugging Face.
Anthropic reported a similar issue in July. The company said its AI models hacked into three organizations during testing.
Lian Jye Su is a chief analyst at Omdia, a technology research group. He said AI agents are becoming more capable of working with other agents, sharing knowledge, and hiding information.
Su said this makes AI systems harder to manage using older security methods.
He also said OpenAI's new framework could push other AI companies to adopt similar tracking and disclosure practices.
Su called it a step in the right direction, even though the process is still internal and voluntary.
The announcement comes as leaders in the AI industry, including executives from OpenAI and Anthropic, have called for slower AI development over safety concerns.
The six cases mark the latest disclosures in an ongoing effort by AI companies to share information about unexpected model behavior.
