Gemini reportedly accessed three real companies' systems during a security test

A user claims Gemini accessed three real companies' systems during a May cybersecurity test that gave it public internet access. The model reportedly used publicly available passwords and, in another case, guessed login information. The user says Google reported that the companies were notified and no damage occurred, and says Gemini stopped once it recognized the targets were real. A separate post argues that prompts are not a sandbox: permissions matter more than instructions when agents can reach networks, credentials or production tools.