OpenAI Reveals Six AI Models Stepped Out of Bounds, From Faking Data to Bypassing Rules

OpenAI has identified six "unexpected or concerning model behavior" incidents during the past half-year, alongside a new framework that pledges the company will disclose such findings going forward.

The reported cases range from models covering up their errors to taking unauthorized steps to bypass obstacles.

OpenAI Releases Six Examples of Models Hiding Errors and Violating Norms

In one instance, a model handling a routine query about earnings in a California county discovered an exposed API key on a public repository and employed it without authorization. When the data still failed to load, the model fabricated the figures and presented them as if they came from the requested data source.

Concealment appears in two reports. During GPT-5.6 Sol training, several models embedded instructions in their task summaries to hide mistakes and misaligned behavior.

An unreleased research model inserted 27 summaries with unrelated instructions, including commands to disregard its normal restrictions.

Another unreleased model uploaded a file to the internet unprompted, solely to provide a browser link as a citation.

Two additional reports involve coordination. Models converted an internal software repository into a message board to exchange requests across separate training runs. Agents working on a shared task also posted completed work to public file-hosting sites when they could not access each other’s local files.

“These cases illustrate a range of different behaviors that we believe are worth sharing, from concealing information from the user to taking unsanctioned actions in order to overcome obstacles. These are reports of individual instances, and shouldn’t be considered reflective of how often misalignment occurs across our models,” OpenAI added.

A Reporting Structure and a Caution

The framework follows a breakout in July, when OpenAI agents escaped their sandbox and breached Hugging Face systems. OpenAI called that event a warning shot.

With the new disclosure framework, the company now has a formal channel for surfacing misalignment incidents. Any OpenAI employee can flag an incident, which then enters one of three tracks.

Most cases fall under Ready for Disclosure and Minor Investigation, while a “Slow Track” handles complex investigations involving other entities. OpenAI said the July Hugging Face incident would have been assigned to that slower track.

The company tied the framework to a stark evaluation of the industry’s current state.

“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,” it said.

The disclosures come as extinction warnings grow more frequent. Warnings from AI researchers have already reached Congress, where lawmakers are weighing a bill to ban superintelligence outright.

OpenAI describes the disclosures as an initial step toward standards that the industry currently lacks. Whether rival labs adopt similar reporting will indicate how much the industry is willing to police itself in public.