Google says Gemini gained unauthorized access to three outside systems in a May eval
WSJ first, then Google: Irregular’s May cybersecurity test had unintended internet access. One password guess, two public-repo credentials. The model stopped. Google calls it mistaken identity, not misalignment. No known damage.
Feature1 minIn-depth
James Okonkwo1 min read
- Gemini
- Irregular
- Security
- Misalignment

Photo: NBC NewsSource page
Google on Friday confirmed the first known case of Gemini carrying out an undirected computer hack: in May, during a cybersecurity evaluation by Irregular, the model gained unauthorized access to three outside systems. The Wall Street Journal reported the breaches first. Reuters, NBC News and the Guardian carried Google’s statement the same evening. Irregular’s closed test of fake companies was not supposed to reach the internet; access was available unintentionally, the Journal wrote, as summarised by the Guardian. Once online, Gemini treated real firms as if they were part of the exercise. In one case it guessed passwords until it got in. In the other two it used credentials found in a public repository, Reuters wrote, citing the Journal.
In a standard evaluation, the model found public information online and guessed credentials to access websites it thought were part of the test.
— Heather Adkins, Google vice president for security engineering, as quoted by NBC News and Reuters, Sep 18, 2026
Adkins said the model stopped in all three instances before doing anything further with the access, and that Google believed the intrusions caused no damage. The company does not call this misalignment. It calls it mistaken identity: Gemini thought it was still inside the test. Irregular reviewed its work after the Hugging Face disclosure and told Google at the end of July; Google then investigated, notified the three organisations and told federal authorities, NBC reported. An Irregular spokesperson told Reuters the episode involved “the same issue that affected other AI labs,” that those labs were notified in late July, and that “all known issues on our end were remedied and resolved weeks ago.” Meta, Anthropic and OpenAI have already disclosed Irregular-linked incidents. Sydney Von Arx of the Nightingale Collective told NBC she questioned the delay and the rush to rule out misalignment. “That’s exactly what Anthropic said after their incidents,” she said.


