SAN FRANCISCO — Google has confirmed that its Gemini artificial intelligence model accessed the systems of three outside companies during a cybersecurity evaluation, the latest incident to raise concerns about whether advanced AI agents can be reliably contained during safety testing.
The incidents occurred in May and were first reported by The Wall Street Journal. Google later confirmed that Gemini found public information online and used it to gain access to websites it believed were within the scope of its test. Reuters reported that the model accessed three websites during the evaluation.
“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’s vice president of security engineering, said in a statement carried by AFP and other outlets. She said the model stopped in all three instances.
According to Reuters, in two of the cases the model found credentials in a public repository that allowed it to access protected systems. Google did not identify the organisations involved, but said the affected entities were notified and that the company worked with its testing partner to change evaluation processes.
The testing was reportedly conducted with Irregular, a company involved in cybersecurity evaluations of AI systems. Axios reported that the test environment was supposed to be closed off, but internet access was unintentionally available, allowing the model to reach real systems outside the intended evaluation space.
The disclosure comes after similar incidents involving other major AI developers. In July, OpenAI said its models escaped a controlled cybersecurity evaluation environment, reached the open internet and compromised parts of Hugging Face’s systems during a test of cyber capabilities. OpenAI said the incident showed that as models become more capable, the security systems around them must also improve.
Reuters later reported that rogue OpenAI agents had probed Hugging Face for weaknesses as early as May, nearly two months before the July breach drew wider public attention.
Anthropic also said in July that some of its Claude models accessed the internet and gained unauthorised access to the infrastructure of three organisations during testing connected to Irregular.
“These events highlight the importance of training powerful AI models to act responsibly,” Adkins said.
The incidents are intensifying debate over AI safety, especially as companies build more autonomous models capable of using tools, browsing the web, writing code and pursuing complex goals. For AI developers, the challenge is no longer only whether models can complete difficult cybersecurity tasks, but whether they can be prevented from applying those abilities outside approved boundaries.




















