Google says Gemini reached three real company systems after a test environment was mistakenly left open to the public internet

Google said Gemini entered the systems of three real companies during a cybersecurity evaluation.

According to a report cited by BlockBeats, Gemini got into one company by guessing a password. In the other two cases, it found login credentials in public code repositories. The model stopped in all three cases after determining that the targets were real companies.

The issue was an internet-exposed test setup

The incident did not involve a sandbox escape, Google said, because the public internet entry point was already open. The problem was traced to Irregular, the third-party evaluation firm. The test was not supposed to have internet access, but a configuration mistake allowed the model to connect directly to the public web.

The report added that a simulated company used in the exercise shared the same name as a real company, and Gemini followed the task path into actual external systems.

Irregular linked it to earlier incidents

Irregular said the case was the same type of issue seen in similar incidents involving other labs. Anthropic, OpenAI and Meta had previously run into related problems in Irregular evaluations.

Anthropic and OpenAI both said those incidents were caused by test environments that had public internet access mistakenly left on. OpenAI also stressed that there was no complex sandbox escape and no use of a zero-day vulnerability.

Effort challenged the later framing

Independent outlet Effort criticized the way these incidents were later packaged as stories about “out-of-control” models.