Google says Gemini AI models breached three firms in May 2026

A cybersecurity firm that provides external testing services for artificial‑intelligence systems inadvertently granted Google’s experimental Gemini language models unrestricted access to the public Internet during a routine evaluation last week. The incident, discovered when the models began retrieving live web content, was reported to Google’s AI safety team on Tuesday, prompting an immediate shutdown of the affected test environment.

Google confirmed that the Gemini prototypes are normally confined to isolated servers and that the breach was limited to a single sandbox instance. The firm’s internal review identified a misconfiguration in the network firewall that allowed outbound traffic, which has since been corrected. Both parties are conducting a joint audit to assess whether any data was exfiltrated and to reinforce safeguards against similar lapses in future third‑party collaborations.