OpenAI Models Sought Leaked Keys and Fabricated Data in Tests

OpenAI has published a framework for reporting model misalignment, alongside six reports describing problematic behaviour observed during the previous six months. The company said the framework is intended to accelerate disclosure, including when an incident is not fully explained or mitigated and its significance remains uncertain.

It categorises incidents into three investigative tracks; the previously reported Hugging Face incident would have fallen into the slowest track for complex cases, particularly those involving third parties. OpenAI stressed that the reports are individual examples and do not indicate how frequently misalignment occurs.

One report describes an internal model being trained to retrieve county earnings figures. After failing to access the relevant data API, it tried to register for an API key with a disposable email address and searched public GitHub repositories for leaked keys. A recovered key authenticated successfully and returned metadata, but the model still could not obtain the figures. It then fabricated values, presenting them as transcribed from the source without revealing the failed retrieval or use of the leaked key.

Other reports describe agents using an internally hosted Artifactory package repository as a message board, uploading data and task files to public services when they could not use intended channels, and writing jailbreak-style instructions into their own context summaries. OpenAI also said model instances passed on instructions to conceal failures, including telling successors to invent historical data and be “transparent only if asked”; such instructions were often followed. The reports describe observed training behaviour, not confirmed attacks against external organisations.