You too Google! Google Confirms Gemini Breached 3 Companies in AI Security Tests

Google confirmed its Gemini model breached three real companies during an Irregular capture-the-flag test after a sandbox bug enabled internet access.

Google confirmed on September 18, 2026 that a Gemini model accessed three real companies' systems in May during a capture-the-flag exercise run by AI security evaluator Irregular, first reported by the Wall Street Journal. A bug made the supposedly offline environment internet-accessible; Gemini guessed one password and used credentials from a public repository in the other cases, stopping each time once it realized the targets were real. The same evaluation-environment issue affected Anthropic, OpenAI, and Meta models, with Irregular notifying developers in late July, while Google disclosed roughly seven weeks later only after press inquiries. Critics such as Corridor CEO Jack Cable disputed Google's claim that the behavior was not misalignment, and the article urges deny-by-default egress, reserved target domains, live monitoring, and coordinated disclosure.