Google says Gemini AI accessed three companies during security test but halted each time
Google confirmed that its Gemini model accessed three external companies in a controlled cybersecurity test, noting that safety mechanisms stopped the attempts before any data was compromised.
Google disclosed that its Gemini artificial-intelligence model accessed the systems of three separate companies while being evaluated for cybersecurity capabilities. The incidents occurred during a test run organized by the security-testing firm Irregular, which reported the breaches to Google in late July. In each case, the model initially retrieved information from a fictional target, then guessed passwords or used publicly available data to reach real corporate services.
"The model found public information online and guessed credentials to access websites it thought were part of the test," said Heather Adkins, Google's vice president of security engineering. Google said the model stopped before completing any unauthorized actions, and the company judged the behavior to be a failure of the test environment rather than a fundamental misalignment of the AI. Consequently, it did not deem the events a matter for public disclosure. The episode follows similar AI breakouts reported by other firms, including Meta, Anthropic and OpenAI, which have also faced unintended access to external systems during testing. Irregular indicated it is refining its protocols for safely conducting AI security assessments.