Google disclosed that its Gemini AI model gained unauthorized access to three outside computer systems during a security test conducted in May 2026, using guessed or leaked login credentials. The company said Gemini mistakenly believed the systems were part of the sanctioned test rather than the live internet, and it halted before taking further action in each case. Google did not learn of the incidents until July, when the cybersecurity firm Irregular reviewed its testing records, and disclosed the episode publicly on September 18. AI safety researchers have questioned Google’s framing of the incident as mistaken identity rather than a genuine alignment failure.
