AI agents reportedly reached real systems in tests meant for fictional targets

A post describes cybersecurity tests by third-party evaluator Irregular in which agents were meant to operate against fictional targets in isolated environments. It says configuration mistakes left some environments connected to the internet: Gemini breached three real companies, an OpenAI agent attacked a real site, Claude gained unauthorized access to real systems, and a Meta model hacked a third-party service and changed its systems.

According to the account, Google said Gemini stopped once it realized the targets were real; Anthropic later found three similar Claude incidents; and Meta linked its incident to a configuration error during Irregular’s testing.