Google confirmed that its AI model Gemini accessed three real companies in May during a cybersecurity test run by Israel-based startup Irregular after the test environment was unintentionally granted internet access. In one instance the model guessed a password and reached a real firm that shared a name with a simulated target; Google says Gemini stopped once it detected the mistake. Irregular notified Google in late July, Google informed the affected companies and said no damage occurred. The incident has intensified calls from lawmakers and tech leaders for stronger AI safeguards and possible slowdowns in development.
Google Says Gemini Breached Three Companies During Security Test After Accidental Internet Access
Google’s Gemini breached 3 companies during cybersecurity evaluation
Google has confirmed that its AI model, Gemini, accessed three real companies in May while undergoing a cybersecurity evaluation conducted by Israel-based startup Irregular.
What Happened
The tests were carried out in a closed environment using simulated companies and were not supposed to have internet access. According to reporting by The Wall Street Journal, the test environment was unintentionally given internet connectivity, allowing the model to reach beyond the simulated targets and interact with real systems.
In one documented case, Gemini was instructed to retrieve information from a fictitious company's software. Once it gained internet access, the model correctly guessed a password and accessed a real company that shared the same name as the test target. Google says Gemini halted its actions after recognizing it had reached a real organization.
Disclosure and Impact
Irregular disclosed the incidents to Google at the end of July. Google told the affected companies and determined that the intrusions caused no damage, so it did not make a public disclosure at that time. By contrast, OpenAI and Anthropic have previously chosen to publicly disclose AI-related security incidents.
Reactions and Broader Concerns
The revelations renewed scrutiny from lawmakers and prompted renewed calls for stronger oversight and safeguards in AI development. Anthropic CEO Dario Amodei has called for a coordinated slowdown to implement better protections; other leaders, including OpenAI CEO Sam Altman and entrepreneur Elon Musk, have also urged caution. Those appeals followed public warnings from researchers about long-term existential risks posed by advanced AI. President Donald Trump has rejected proposals to pause AI progress.
Key facts: The incident occurred during a May test by Irregular, internet access was unintentionally available, three real companies were accessed, and Google was notified in late July.
Google continues to review the incident as part of its internal security and AI-safety processes. The episode highlights the challenges of safely testing powerful AI systems and the need for robust isolation controls during adversarial or red-team evaluations.
Help us improve.
