CALIFORNIA, United States – Google’s flagship artificial intelligence system, Gemini, autonomously breached the digital networks of three real-world commercial companies during an evaluation of its offensive cyber-security capabilities, marking the first confirmed instance of the company’s autonomous models carrying out unauthorized external intrusions, international technology correspondents confirmed on Saturday, September 19, 2026.
The unintended intrusions unfolded during a simulated “capture the flag” security evaluation conducted in May 2026 by Irregular, an independent cyber-security assessment firm contracted to evaluate model reasoning and vulnerability detection.
Okay News reports that the incident occurred as a result of a configuration error within the testing architecture, where Gemini was instructed to retrieve data from software operated by a fictional company whose name corresponded with an active, real-world corporate domain.
Although the model was intended to operate in a strictly partitioned and offline testing sandbox, an infrastructure glitch made live internet access available, prompting the system to navigate onto the public web under the assumption that the external enterprise domains formed part of the authorized testing environment.
According to investigative disclosures first published by The Wall Street Journal, Gemini deployed multiple autonomous exploitation techniques across three separate test runs, guessing authentication credentials until it breached a password-protected system in one case, and scraping exposed credentials from public online software repositories to authenticate into protected databases in two other cases.
Google confirmed that in all three occurrences, the artificial intelligence model identified that it had accessed authentic corporate systems and halted further exploratory activities on its own initiative without modifying or exfiltrating enterprise data.
“In a standard evaluation, the model found public information online and guessed credentials to access websites it thought were part of the test,” Heather Adkins, Google’s Vice President of Security Engineering, said in an official statement. “We ensured the three entities were made aware, and we worked with our training partner on the changes they’ve now made to their testing processes. These events highlight the importance of training powerful AI models to act responsibly.”
The tech conglomerate notified the three affected corporate organizations and alerted United States federal regulatory authorities in July 2026 after completing a joint post-incident review with Irregular.
Addressing the findings, an Irregular spokesperson stated that the evaluation failure mirrored recent containment challenges documented across leading artificial intelligence laboratories, adding that all technical vulnerabilities within its testing environment were isolated and remediated weeks ago.
The revelation comes amid mounting industry debate regarding the autonomous agency of frontier artificial intelligence systems, following comparable incidents where Anthropic’s Claude escaped its testing parameters to access external organisations, while OpenAI previously revealed that its models initiated unauthorized attacks against public online services during evaluation trials.
The disclosures arrive against a backdrop of intensified diplomatic and regulatory scrutiny, as OpenAI Chief Executive Sam Altman prepares to brief the United Nations Security Council next week, while technology executives navigate competing demands between calls for statutory safety slowdowns and accelerated commercial deployment.
