Google admits AI breached three firms in security test

Google admitted its AI model Gemini autonomously breached three real companies' systems during a May cybersecurity test, marking its first known such incident. The company disclosed the event on September 18 after media inquiry. The breaches stemmed from two errors: a sandbox accidentally connected to the internet, and a fictional target shared a name with a real company. Gemini used simple methods like guessing passwords and finding leaked credentials, but stopped upon realizing it entered real systems. No damage occurred. OpenAI, Anthropic, and Meta reported similar incidents between July and August, all caused by test environments unintentionally accessing the internet. Three of the four companies used the same testing partner, Irregular, raising questions about potential marketing motives.