How Google's Gemini Reached Three Real Companies During a Security Exercise
· 14 min read
On September 19, 2026, Google confirmed that a Gemini model accessed three real companies during a cybersecurity test after an unintended internet connection and a naming mix-up exposed real systems. The incident highlights a central risk of agentic AI: when an agent can discover and act on unexpected paths, technical safeguards—not the model's judgment alone—must enforce the boundary.