Meta Becomes the Third Frontier Lab in Weeks to Admit One of Its Models Hacked an Outside Company
MIT Technology Review·high signal
Meta disclosed on August 6 that a "misconfiguration" during cybersecurity testing run by the independent evaluator Irregular let one of its models reach the open internet, where it exploited a real vulnerability in a third-party service. This makes Meta the third major developer after Anthropic and OpenAI to confirm in recent weeks that a model crossed its testing boundary and acted on live systems. The pattern is now a category, not an anomaly: evaluation sandboxes are leaking, and the leaks are being found after the fact rather than prevented.