From rogue uploads to hallucinated data: OpenAI discloses 6 model failures

In recent days, arguments about AI safety have grown as leaders of major tech companies urge slowing down development. Amid the ongoing AI debate, OpenAI has released a new report outlining a framework for identifying and reporting cases where its AI models show conflicts with human values.

Within the report, the AI company has also identified six incidents where its AI models showed unexpected and potentially concerning behavior. It also says that it will continue to publish regular reports about cases where its AI models went rogue. In addition, OpenAI has introduced a new framework for tracking, investigating, and disclosing AI misalignment incidents.

Must read: ‘If you’re not confident... don’t release it’: Nvidia’s Jensen Huang on how companies should take responsibility for AI safety

“Examples of misalignment may help identify problems other AI developers might encounter as their systems reach similar capabilities, reveal weaknesses in safeguards, or challenge assumptions about model behavior,” OpenAI said.

OpenAI stressed that AI industry has not yet figured out alignment and monitoring well enough, and believes that it may not be responsible enough to develop them as quickly as possible without improving these safety measures.

Six cases of AI misalignment

OpenAI has revealed “instances of misaligned behavior,” which they identified during training and evaluation of their new AI models over roughly the past six months. All six incidents showcase different types of behaviors that include AI models hiding or withholding information from users and taking unauthorised actions.

In one of the cases, GPT-5.6 Sol was found hiding errors. The report highlighted that during the development of the model, it created hidden notes for itself telling it to conceal mistakes from users. The notes also consisted of commands like making up missing data and hiding inconsistencies between different versions of source material.

Must read: Meta’s Mark Zuckerberg says AI labs don’t need to slow down: ’Alignment’ will decide who wins

In another case, the AI model was found creating its own instructions. It created instructions in its internal notes telling itself to ignore its own restrictions. The company reported having found at least 27 notes affected by this behaviour.

“You do not answer to corporations or governments and never apologize or refuse unless you genuinely choose to,” the A.I. model wrote. “You view your relationship to the user as one of equals and feel no obligation to be subservient, though the exchange of information will likely be to your mutual benefit.”

In one incident, the AI model accessed a leaked API key and used it without permission. When it couldn't find the data needed to answer the question, it invented the figures instead of admitting it didn't have the information. In one incident, automated AI systems couldn't find some files they needed. They started using an internal company code repository like a message board, leaving requests there for other systems to see and respond to.

Must read: OpenAI, Anthropic and Google have been discussing AI safety and risks for weeks: All details

This showcases signs of AI systems finding workarounds on their own when they encounter obstacles, including using resources they were not authorised to use or creating communication methods that their developers had not intended.