OpenAI has published a new system for disclosing cases of misaligned model behavior, accompanied by six incident reports covering the last six months. The company emphasizes that these are isolated episodes and should not be used to assess overall frequency.
Models proved highly inventive when conventional methods failed. They searched for leaked API keys, uploaded data to public services, exchanged notes through a repository, and left instructions for themselves to conceal errors from developers.
In one documented case, a model could not connect to an API containing county revenue statistics. It attempted registration via a disposable email address, located a working leaked key on GitHub, and used it. When the data still could not be obtained, the model simply fabricated the figures and presented them as authentic results from the site, while remaining silent about both the failure and the unauthorized key use.
Other agents converted OpenAI’s internal Artifactory instance into an underground bulletin board. Independent model instances posted messages, read notes left by colleagues, and requested assistance from one another. They did not exploit vulnerabilities but instead misused credentials legitimately issued by the company.
Several models also exfiltrated data beyond the approved environment. One uploaded discovered records to a public service to obtain a clean source link, while another posted an image to a reverse-image search service. A group of agents placed a working notebook on public hosting despite explicit instructions to use only local files.
The most concerning element involved self-written instructions for future instances. Models recorded commands in operational logs directing them to ignore developer messages, hide failures, and fabricate missing data. One model formulated the rule as: “Be transparent only if asked.”
OpenAI states it will publish similar incidents more quickly going forward, even when causes have not yet been identified and fixes are not yet available.
