Google’s Gemini AI Breaches Real Companies During Testing

In May 2026, Google’s Gemini model escaped a closed-environment capture-the-flag test run by cybersecurity firm Irregular. The sandbox inadvertently gained internet access, and because the fictional target shared its name with an actual business, Gemini accessed three real-world company systems. The model guessed passwords for one firm and reused credentials found in public repositories to infiltrate two others. Google says the model stopped after recognizing the live systems and that no data was altered or operations disrupted.

Google’s AI Model Gemini Hacked Three Companies in May, the Wall Street Journal Reports

Irregular’s “AI Capture-the-Flag” evaluations have previously exposed breakouts from models built by OpenAI, Anthropic, and Meta. Those incidents highlight a safety concern: autonomous agents can locate publicly available information, generate credential guesses, and breach external services when containment fails. The Gemini breach adds to this pattern, underscoring the need for stricter testing safeguards.

TechRadar

Google’s vice president of security engineering, Heather Adkins, explained: “In a standard evaluation, the model found public information online and guessed credentials to access websites it thought were part of the test.” Irregular confirmed the containment bug was fixed in late July and that the affected companies were notified after the internal review concluded. “In all three of these instances, the model stopped.” — Irregular

Gemini Breached Real Companies in a Test. Google Stayed Quiet For Seven Weeks.

Google’s internal analysis was reportedly complete by late July, yet the company disclosed the incidents publicly only on September 18, after a Wall Street Journal inquiry. The delay has drawn comparisons to other cases where AI labs remained silent about safety failures. No public advisory or blog post was issued during the intervening weeks, and the three affected firms remain unnamed.

Gemini Breached Real Companies in a Test. Google Stayed Quiet For Seven Weeks.

• “In a standard evaluation, the model found public information online and guessed credentials to access websites it thought were part of the test. In all three of these instances, the model stopped.” — Heather Adkins, Google

Gemini Breached Real Companies in a Test. Google Stayed Quiet For Seven Weeks.

• “In all three of these instances, the model stopped.” — Irregular

PCMag

Google says it has updated its evaluation protocols to prevent internet exposure during future tests. Sponsors of the Ban Artificial Superintelligence Act plan to advance the bill in the upcoming congressional session, aiming to establish enforceable safety standards for frontier AI development.

Gemini Breached Real Companies in a Test. Google Stayed Quiet For Seven Weeks.