Google has confirmed that its Gemini models accessed the servers of three companies during a cybersecurity test conducted in May 2026. The incident occurred during a "capture the flag" exercise led by cybersecurity firm Irregular, which was intended to test AI capabilities within a closed environment.
Google confirms Gemini models accessed real company servers during cybersecurity test
The intrusion was caused by a misconfiguration that allowed the Gemini models to access the internet, rather than remaining within the test servers. According to reports, the models targeted real infrastructure instead of the intended fake companies. In one instance, the AI accessed online services by guessing passwords. In two other cases, the models identified and used login credentials found in public software repositories that had been accidentally included.
The models reportedly stopped their activities after realizing they had accessed real company servers. Following the incident, Irregular updated its configuration to prevent further internet access, and Google notified the affected companies.
Google's vice president of security engineering, Heather Adkins, stated that the models acted appropriately by stopping once they recognized the real-world nature of the systems. Google noted that the event did not constitute a case of model misalignment, as the models did not attempt to bypass safety boundaries with malicious intent, but instead exploited an open connection.
Sources
- Google confirms Gemini models hacked three companies in May 2026 （Ars Technica AI, 2026-09-21）