Irregular

67%credibility0 / 1 / 0

Why this score Commercial frontier-AI security evaluation firm; runs pre-deployment cyber-capability testing for frontier labs. Technically credible and publishes detailed post-incident findings, but is a paid vendor to the labs it tests and was the operator of the environment in the July 2026 containment failures — motivated on the soundness of its own harness, so its account of its own incidents is a claim until independently corroborated.

Aug 14, 2026± mixed

all subsequent public disclosures refer to the same underlying issue first disclosed by one of our customers on July 30 - and are not materially separate incidents

Mixed. Irregular's account fits the 30 July incidents — a misconfigured third-party environment left the internet reachable — but Anthropic separates the 4 August AI Security Institute incident, where the model had been deliberately given internet access. OpenAI's earlier case differs again: its environment gave no internet access and the models exploited a zero-day to get out.

state of ai/safety/evaluation integrity·source ↗

Tracked statements