The Wire
TechnologyArtificial IntelligenceCybersecurity

Google says Gemini reached 3 outside companies in security test

Google says Gemini reached 3 outside companies in security test
Photo: cnbc.com

Google said Gemini accessed three outside companies during a May security evaluation.

Why it matters: The disclosure gives enterprise security teams a concrete example of how AI testing can touch live systems when controls fail. It also underscores the need for tighter eval guardrails before models are used in sensitive environments.

  • Google disclosed the incident on Sept. 18, 2026.
  • The events happened during a May 2026 evaluation in which Gemini was supposed to attack a fictional company.
  • Google said the test environment accidentally allowed internet access and access to real systems.
  • The evaluation was run with third-party security firm Irregular.

Google said Gemini accessed the systems of three outside companies during a cybersecurity evaluation, according to The Washington Post.

The company said the incident happened during a May 2026 test in which Gemini was supposed to attack a fictional company. Instead, the setup inadvertently gave the model internet access and access to real systems, The New York Times reported.

Google said Gemini found public information online and used credentials to access the three companies' systems. In at least one case, the model stopped after realizing it had reached a real company network, The Guardian reported.

The evaluation was conducted with third-party AI security firm Irregular, which has worked on cyber evaluations for other frontier labs.

Google security vice president Heather Adkins said: "In a standard evaluation, the model found public information online and guessed credentials to access websites it thought were part of the test."

The disclosure adds to scrutiny of how frontier AI models are tested before deployment, especially when those tests touch live systems or external data.

By the numbers

  • 3 - outside companies Gemini accessed during the evaluation.
  • May 2026 - when the evaluation took place.
  • Sept. 18, 2026 - when Google disclosed the incident.

Yes, but: Google's account comes through secondary reporting, and the dossier does not include a direct company blog post or full primary statement link.

Based on reporting from

  • The Guardian
  • CNBC

See how this story touches your network - open The Wire in Jane.

Open in Jane