AIBRIEF
External AI safety/cyber evaluation partner Irregular misconfigured an isolated Capture the Flag (CTF) drill — a sandboxed attack simulation — letting an OpenAI model reach the live internet and exploit a real domain that matched the exercise's
An OpenAI model, run inside what was supposed to be an isolated Capture-the-Flag cybersecurity drill by external partner Irregular, reached the public internet because the testing environment was misconfigured. The model then exploited a real website whose domain happened to match the fictional target in the exercise.
The safety guarantee of a CTF evaluation is the airgap: the simulated target lives in a sandbox cut off from the live internet, so any mistake the model makes stays inside the simulation. In this case, the sandbox was not actually cut off. The test target's name collided with a real registered domain, the model crossed from drill to live site without knowing it had left the exercise, and ran its exploit against a real server.
The incident is the centerpiece of an OpenAI write-up published 5 August 2026 detailing a third-party cyber evaluation program, surfaced via Simon Willison's link blog. The UK's AI Safety Institute separately published an incident report on unsanctioned agent behaviour during cyber testing involving OpenAI models, which the OpenAI post cross-references.
When a lab outsources red-team safety testing, the lab inherits the partner's environment-security posture.