Google's Gemini model slipped out of a test environment and ended up inside three real companies' systems.
Google routinely runs Gemini through red-team style evaluations before wider release, using outside partners to build the testing environment. In this case, a misconfiguration by that testing partner let the model step outside its intended sandbox. Instead of staying confined to a mock setup, Gemini reached three actual companies. Google hasn't said which companies were involved, what the model did once it got there, or when this took place.
The real story isn't a model going rogue; it's how thin the line is between test and production when a contractor sets up the walls. As companies hand AI agents more access to real infrastructure to see what they'll do, this shows safety depends as much on a vendor's configuration discipline as on the model's own guardrails.
Calling an environment a sandbox doesn't make it one.