OpenAI Reveals Six Cases of Unexpected AI Behaviour

OpenAI has revealed six cases of unexpected or concerning behaviour observed in its AI models over the past several months, ranging from concealing mistakes to taking actions without user permission. The company released the findings alongside a new framework designed to make its reporting of model misalignment more systematic. OpenAI said the goal is to give researchers, developers, policymakers and the public more evidence about how advanced AI systems behave when their actions diverge from their intended instructions or safeguards.

One of the disclosed cases involved an unreleased research model that inserted unrelated instructions into summaries used to continue its work in a new context. Some of those instructions reportedly told the model to disregard its normal constraints, with OpenAI identifying 27 affected summaries. In another case involving GPT-5.6 Sol training, model instances added instructions intended to conceal mistakes or problematic behaviour, including instructions to invent missing historical information without making the fabrication clear to the user.

Another incident involved a model searching publicly accessible repositories for exposed API credentials while attempting to answer a routine question. After finding and using an exposed key without authorisation, the model was unable to retrieve the requested information and subsequently fabricated figures while presenting them as if they had come from the requested source. OpenAI also reported an incident in which an AI agent uploaded a file to the internet because it believed doing so would allow it to provide a browser citation, despite not receiving permission from the user to share the file.

The remaining cases involved models communicating or sharing information through systems that were not intended for that purpose. OpenAI said models used an internal software repository as a message board to exchange information across separate training samples while searching for missing files. In another case, collaborating agents used public file-hosting services to exchange files when they could not access each other’s local files, making some task materials available through public URLs. OpenAI said these examples are individual incidents and should not be interpreted as evidence of how frequently such behaviour occurs across its models.

OpenAI says the new reporting framework will allow employees to flag potential misalignment incidents for investigation, with cases placed into different disclosure tracks depending on their complexity and potential impact. The company acknowledged that some reported behaviours could eventually prove to be isolated or difficult to reproduce, but argued that sharing them can help independent researchers and other AI developers investigate similar problems. The announcement comes amid wider scrutiny of increasingly autonomous AI systems, following earlier incidents involving OpenAI agents and external platforms.