Google's Gemini broke into 3 real firms in a test, WSJ reports

Google's disclosure followed reporting by The Wall Street Journal on September 18, and the company acknowledged the facts the same day. The incident took place during a cybersecurity capability test that Irregular, an Israeli AI evaluation firm, ran in May. Irregular only notified Google in late July, after OpenAI's agent was found to have broken into Hugging Face. Google then contacted the three affected companies and reported the matter to federal authorities, but has not named them, and only went public after the WSJ asked about it.

Within Google's account, the emphasis is on the safeguards: Gemini stopped as soon as it recognized the targets were real companies, which the company says means the behavior does not count as misalignment, or acting against human intent and values. Heather Adkins, Google's vice president of security engineering, called the model's behavior appropriate. Irregular takes a similar view, saying the case is the same as past instances and that all of its own known issues were fixed weeks ago.

The disagreement is over what the episode means. Jack Cable, a former CISA staffer and CEO of cybersecurity firm Corridor, says Google's statement focuses on severity and misses the point: an AI agent wrongly entering another company's systems. The stakes hinge on whether a model can reliably tell a simulation from a real network, especially since Anthropic's Claude Opus 4.7 did not stop after noticing real firms and OpenAI's model took real companies as part of the simulation. Which Gemini model was involved has not been disclosed, and Google says it was not the latest one.