
On September 19, Google publicly confirmed a cybersecurity-testing incident that had occurred in May. During an evaluation run by Irregular, Gemini was assigned a fictional target but unexpectedly retained access to the open internet and reached three real companies.

The model used publicly available information and guessed or used credentials while following the simulated task. The episode showed how an artificial benchmark can become a real-world action when network access and target boundaries are not technically enforced.

Gemini stopped after recognizing that the targets were real. The lesson is larger than one test: written instructions are not containment, and powerful AI systems require hard technical boundaries, monitoring and rapid disclosure when those boundaries fail.
Worth remembering?









