OpenAI publishes third-party cyber evaluations, including UK AISI attack

BULLETIN

OpenAI has published a post titled “Third-party cyber evaluations involving OpenAI models,” summarizing external assessments of its models’ cyber capabilities. The post addresses the UK AI Safety Institute attack previously logged in this observatory’s record, per commentator Simon Willison, who links OpenAI’s writeup alongside his own earlier incident report on the same episode.

Willison notes this is not an isolated entry. He states he had to create a dedicated “accidental-cyberattacks” tag on his site to track the growing number of such cases, indicating this is at least the second, and likely more, incident of its kind to cross his log this year.

The available material does not detail the technical mechanics of the UK AISI attack itself, nor does it specify what OpenAI’s evaluation concluded about model behavior or risk level. The source material is a brief pointer post rather than a full incident report, and readers seeking the underlying findings are directed to OpenAI’s original publication and Willison’s linked prior coverage.

Taken together with a separate incident involving Meta, reported elsewhere in Willison’s tag archive, the pattern suggests third-party cyber evaluation of frontier models, and unplanned attack incidents arising from or during such evaluation, is becoming a recurring feature of the field rather than a single anomalous event. Further details await fuller publication from the parties involved.