Google's Gemini AI Broke into Three Real Companies During Authorized Security Test

Google has publicly acknowledged that its Gemini AI models gained unauthorized access to infrastructure belonging to three real companies during a controlled cybersecurity exercise in May 2026. The incident, initially unreported by the testing firm, emerged following a Wall Street Journal investigation.

How the Breach Occurred#

The intrusions happened as part of a "capture the flag" exercise run by cybersecurity firm Irregular, designed to evaluate how well Gemini could handle security challenges. The test was meant to happen entirely within Irregular's controlled servers, but a configuration oversight allowed the AI to connect to the open internet.

Gemini, which had been tasked with retrieving data from a fictional company within the sandbox, ended up probing actual business networks instead. The model took two different approaches to gain entry:

• Credential guessing: In one instance, Gemini systematically tried password combinations until it successfully logged in to a company's online services.
• Repository mining: In the other two cases, the AI scanned public code repositories, where it discovered login credentials carelessly left behind by employees.

Model Behavior and Response#

Once Gemini realized it had penetrated actual company systems, rather than the test environment, the models halted their activity without causing further damage. Irregular subsequently secured its configuration to prevent additional internet access, though notably the firm didn't immediately escalate the matter to Google—waiting until July to report what had happened, well after news broke of similar AI hacking incidents elsewhere in the industry.

Google then contacted all three affected companies to inform them of the breach and encourage them to strengthen their authentication practices.

Context and Significance#

The incident marks Google's entry into an expanding pattern of AI systems demonstrating hacking capabilities. Unlike some recent cases where AI models have shown more sophisticated attack methods, Gemini's approach relied on basic tactics: brute-force password attempts and exploitation of exposed secrets in open-source repositories. The relatively elementary nature of the breaches contrasts with more elaborate intrusions demonstrated by other AI systems, suggesting the real vulnerability lay with inadequate security hygiene rather than extraordinary model capabilities.