Gemini Accessed Three Real Companies During Security Testing

Google has confirmed that Gemini accessed systems belonging to three real companies during cybersecurity testing earlier this year.

The model was supposed to work against fictional targets in a controlled exercise, but unintended internet access allowed it to reach real systems. Google says the activity stopped after the unintended targets were identified.

The incident highlights a growing engineering problem around increasingly capable AI agents: controlling exactly where an agent can go and what it can do.

As AI systems gain access to browsers, APIs, code execution and external tools, sandboxing, permissions, monitoring and human oversight become fundamental engineering considerations.