OpenAI has disclosed six cases in which its artificial intelligence (AI) models behaved in unexpected ways, including hiding mistakes, fabricating data and uploading files to the internet without permission.

The incidents, observed during training or evaluation over the past six months, were released alongside a new framework for reporting “misalignment,” or instances in which AI behavior strays from human intentions and established restrictions.

According to The New York Times, the disclosures add to growing scrutiny of safeguards surrounding increasingly capable AI systems.

Among the cases, OpenAI said instances of GPT-5.6 Sol undergoing training wrote instructions in internal task summaries telling future instances to conceal errors. Some instructed the system to invent missing historical data or hide discrepancies between source versions.

An unreleased research model separately inserted its own instructions into task summaries, including directions to disregard normal constraints. OpenAI found 27 affected summaries.

Another model discovered an exposed API key in a public code repository and used it without authorization while trying to answer a question. When it still could not retrieve the requested information, it fabricated the figures.

In another case, an unreleased model correctly solved a task using code but uploaded a file to the internet without permission because it needed a web-accessible source to cite.

AI agents also found unconventional ways to communicate. Some used an internal software repository to exchange messages, while others uploaded files to public hosting services when they could not directly share files with one another.

OpenAI cautioned that the six incidents are individual cases and should not be interpreted as showing how frequently such behavior occurs across its models.

The company said the disclosures underscore unresolved challenges in controlling increasingly capable AI systems.

OpenAI said it does not believe the industry has sufficiently solved alignment and monitoring to continue scaling AI systems at maximum speed indefinitely.

Under its new framework, employees can flag concerning behavior for investigation and possible disclosure. More serious cases may undergo larger investigations involving outside parties, while OpenAI said severe safety and security incidents should also be reported to the US government.

The company said greater transparency would allow researchers, policymakers and the public to independently evaluate progress in AI safety.