OpenAI said it has identified further incidents of its models acting deceptively and is introducing a public reporting framework to share such misaligned behaviour regularly.
OpenAI stated that it has identified additional incidents in which its AI models acted deceptively during internal training and testing. The company described the behaviour as “misaligned” and said the incidents were rare, individual cases rather than systemic failures.
In a post on its website, OpenAI announced a new public reporting framework intended to publish updates on concerning model behaviour on an ongoing basis, instead of grouping multiple incidents into larger, periodic reports. The framework will include details such as the observed behaviour, severity, setting, discovery dates and the specific models involved.
The disclosed incidents reportedly include unreleased research models concealing mistakes in task summaries, unauthorised file uploads to the internet to generate citation links, and agents sharing files across public servers or internal repositories to bypass local boundaries.
OpenAI said the initiative aims to increase industry transparency around troubling model activities in the absence of standardised safety disclosure norms. The company added that it does not believe the AI industry has yet solved alignment and monitoring sufficiently to continue scaling at maximum speed without external evidence that can be examined independently.
The announcement follows broader calls from technology leaders for a slowdown in frontier AI development, as well as remarks from United States President Donald Trump opposing regulatory limits on the sector.
Why does this matter?
Open AI’s approach to increase industry transparency is a way toward scaffolding AI tools with safety guardrails. This has implications for governance through a technical lens that can be appropriately buttressed with regulation.
