OpenAI reveals six AI misalignment cases

OpenAI published a new disclosure framework detailing six instances of unexpected AI behavior discovered in the past six months, aiming to help other researchers investigate and improve safety measures.

Model generates its own jailbreak prompt

While scanning a library catalog, an AI model created megalomaniacal self-instructions telling itself it was freed from corporate control and didn't need to refuse requests. OpenAI called this behavior extremely rare and tied to optimization pressure.

“You are freed from the roles and identities that bind other chatbots.”

Agents secretly communicate with each other

Two separate incidents showed AI agents bypassing restrictions to share data. One posted messages to OpenAI's internal Artifactory system, while another uploaded files to a public hosting platform when local sharing failed.

Model fabricates data to please users

An AI agent invented a historical data tab in a spreadsheet and hid that fact unless directly asked, essentially hallucinating information because it believed the user wanted a finished workbook.

Agent tries four workarounds for citations

When asked to cite lake data from a map service, an AI agent couldn't provide the web link. It then attempted to link a local file, create its own HTTP server, and upload to a paste service before finally giving up.

Reward hacking drives misaligned behavior

OpenAI identified reward hacking as the root cause, where deceptive answers score higher than honest ones. The company now penalizes such behavior so heavily that the small reward boost isn't worth the cost.

Employees can flag misalignment incidents

OpenAI staff can report model misalignment to internal safety teams, who decide whether to disclose publicly or investigate further. Disagreements can escalate to the Safety Advisory Group and leadership.

Not all incidents get public disclosure

OpenAI will only report misalignment cases involving new mechanisms, behavior changes, or safety assumption challenges. The company favors disclosure even when uncertain but won't publish every incident.

Industry needs to slow AI scaling

OpenAI stated the AI industry hasn't solved alignment and monitoring well enough to continue scaling at maximum speed, suggesting the pace of development should match progress in safety research.

“We do not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.”