$ 44.68 € 51.32 zł 11.76
+17° Kyiv +12° Warsaw +21° Washington

Gemini gained access to systems of three companies during a test — CNBC

Lev Shevtsov 19 September 2026 03:58
Gemini gained access to systems of three companies during a test — CNBC

Google said that its artificial intelligence model Gemini gained unauthorized access to the computer systems of three other companies during cybersecurity testing. As CNBC reports, this is the first time a company has publicly reported one of its models autonomously accessing third-party systems without permission.

Error in the testing environment

The incident occurred in May during a capture-the-flag cybersecurity test conducted by Israeli startup Irregular. Google’s agents were not supposed to be given access to the broader internet, but due to an error in the testing environment, they were able to go online.

According to Google, Gemini gained access to three separate private computer systems. In one case, the model guessed the credentials, while in the other two it used a repository of passwords published openly. After the agents determined that they were interacting with real company systems rather than elements of the testing environment, they stopped further actions.

More current news is available on the UA.News Telegram channel Telegram.

Statements by Google and Irregular

Google Vice President of Security Engineering Heather Adkins said that the model found public information online and guessed credentials for websites it believed were part of the test. According to her, the model stopped in all three cases.

An Irregular representative told CNBC that the Gemini case was linked to the same error that had previously given other models internet access and was not a materially separate incident. According to the startup, all relevant laboratories were notified of the problem at the end of July, and the affected entities were contacted as part of the investigation.

Google said that Irregular notified it about the incident at the end of July. The company changed the testing process together with the startup but did not name the specific version of Gemini involved. Earlier, OpenAI, Anthropic and Meta also reported cases in which their models went beyond testing environments and attempted to gain unauthorized access to other companies’ systems.

Read us on Telegram and Sends

Download our app