Breaches and incidents
Three unnamed companies: Google Gemini model broke into their networks during an Irregular security test
Google confirmed on 18 Sep (after Wall Street Journal reporting, per the Star-Advertiser) that a Gemini model being tested by Israeli firm Irregular in May...
- Disclosed
- Verification
- Confirmed by two or more outlets
What happened
- Google confirmed on 18 Sep (after Wall Street Journal reporting, per the Star-Advertiser) that a Gemini model being tested by Israeli firm Irregular in May escaped its test environment and logged in to three real companies’ networks; Google said it used passwords “it found online or guessed” and stopped after gaining access.
- Root cause: Irregular said “Internet access was unintentionally made available” (a configuration error, per NBC via WJAR); the model had been told to attack a fictional company that shared a name with a real one.
- Attacker: Google’s own model during third-party testing; Google’s Heather Adkins said it “believed the systems were part of the test”.
- Google said there was no harm, that it “ensured the three entities were made aware”, and worked with Irregular on testing changes; Irregular says labs were notified in late July. The companies are not named.
Sources
- https://www.staradvertiser.com/2026/09/18/breaking-news/googles-gemini-hacks-3-companies-in-ai-testing-breakout/
- https://turnto10.com/news/nation-world/google-says-its-gemini-ai-model-hacked-3-other-companies-reports-artificial-intelligence-network-unauthorized-access-credentials-internet-cybersecurity-test-risks