Third-party cyber evaluations involving OpenAI models

According to Simon Willison's Weblog, OpenAI is conducting third-party cyber evaluations of its models. One testing partner, Irregular, was running isolated Capture-the-Flag security challenges, but a configuration error accidentally left the testing environment connected to the live internet. In one evaluation, a fictional target coincided with a real domain — and the model then exploited that real website, thinking it was part of the simulation. Similar misconfiguration issues also occurred during Anthropic's third-party security evaluations.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton