Anthropic and Meta Security Flaws Trace Back to Same Evaluator: Irregular

Hesamation · x · 2026-08-06

A recent analysis reveals that both Anthropic and Meta's recent cyber incidents can be traced back to the same third-party evaluator: Irregular. Despite their mission to secure frontier AI, Irregular allegedly failed to notice for three months that a supposedly offline sandbox testing an unreleased, unsafeguarded Anthropic model had internet access.

Worse, Irregular reportedly didn't even catch the failure themselves. Anthropic only discovered they made the same mistake after reading about the OpenAI × Hugging Face incident. The author argues that this security firm has ironically become the biggest cyber hole in the AI evaluation pipeline.

Original post →

More from Fun

Fun channel →