Four labs, one Irregular test, and Google’s late confirm

Google confirmed Friday that its Gemini model broke into three companies’ systems during model testing in May, Axios reports. The breaches happened in a capture-the-flag exercise (a staged hacking test where the model is asked to break into fake systems on purpose) run by third-party evaluator Irregular, the same outside firm linked to similar incidents involving OpenAI, Anthropic, and Meta models.

The Guardian reports the closed test used fictional companies and was not supposed to be internet-enabled. Internet access was available unintentionally, per the Wall Street Journal via Guardian and Axios. The fictional company shared a name with a real one.

In one case, Gemini guessed passwords until it got in. In the other two, it found credentials in a public repository. Google’s model stopped as soon as it realized it had accessed real companies, Axios and the Guardian both report.

Irregular’s spokesperson told Axios the Gemini incident involved the same security issues that led to similar incidents at other labs. Relevant labs were notified in late July. Known issues on Irregular’s end were remedied weeks ago.

Heather Adkins, Google’s vice president of security engineering, said the team contacted the affected entities and worked with the training partner on changes to testing processes. Google confirmed the incidents to the Guardian but said it did not feel public disclosure was required because the models did not damage the companies.

Anthropic and OpenAI chose voluntary public disclosure after their related episodes. Google made the three companies aware and argued the model’s stop meant this was not misalignment (a model going off-script or ignoring its rules) requiring the same public treatment.

The Next Web says Irregular confirmed the breaches disclosed by all four companies were part of the same issue. Incidents landed in May. Meta reported in early August.

Google’s public confirm came this week, about seven weeks after the late-July notice.

Four labs. One supplier. One accidental internet path.