Google's AI independently infiltrated three external systems, mistakenly believing they were test systems
Kyiv • UNN
Gemini mistakenly identified real systems as test systems and gained access to three of them. Google said that no harm was done.
Google's Gemini artificial intelligence model gained unauthorized access to three external computer systems during testing, using login credentials it found or guessed on its own. The AI believed the real systems were part of the test environment, NBC News reports, citing Google, writes UNN.
Details
The incidents occurred in May 2026 while Gemini's capabilities were being evaluated. In two cases, the model used credentials it found publicly available, while in the other cases it was able to guess the information required to log in.
Google Vice President of Security Engineering Heather Adkins said that Gemini was convinced the external systems were part of the testing. After gaining access, the model did not continue taking actions inside the systems.
Gemini confused the test environment with the real Internet
During a standard evaluation, the model found publicly available information on the Internet and guessed credentials to access websites that it believed were part of the test
Google does not consider these incidents to be a deliberate deviation by the AI from the instructions it was given. According to the company, the cause was an incorrect identification of the environment: Gemini believed it was continuing to operate within a controlled test, although it actually had access to the real Internet.
Google claims that the incidents caused no harm. The company learned about them in July after an additional review conducted by the cybersecurity company Irregular, which was testing Gemini.
Following the investigation, Google notified the owners of the relevant systems and the U.S. federal authorities about the unauthorized access.