Google’s Gemini AI Model Hacked Three Companies’ Systems During Security Check
Mian Nadeem
Published: September 20, 2026 | 06:36 PM
This was the second such incident, while the first case of Gemini accessing a real system outside a testing environment occurred in May, as reported by The Wall Street Journal.
Google Confirms AI Model Hacking
Google has confirmed that its artificial intelligence program, Gemini, hacked the systems of three companies during security checks, although it stopped itself before completing the actions each time.
According to a report on September 19, 2026, Google admitted that its AI model Gemini breached the security of three companies during a test designed to assess its cybersecurity capabilities. The Wall Street Journal stated that:
- The initial incident happened in May when a cybersecurity firm, "Irregular," was conducting a test.
- During the test, Gemini was tasked with obtaining information about a fictional company but gained inappropriate access to the internet.
In the first instance, the model successfully guessed a password and accessed a real company's service.
Heather Adkins, Google’s vice president of security engineering, explained to Al Jazeera that in other cases, the model searched for publicly available data online and guessed credentials, which could have provided access to websites it believed were part of the test.
The company revealed that this occurred three times, and each time, the model halted itself before completing the action. Irregular informed Google about these hacking incidents at the end of July.
Google stated that:
- This behavior was not an example of model misalignment or going in the wrong direction.
- There was no need to disclose it publicly as Gemini’s safety measures had been effective.
Similar incidents have been reported by Meta, Anthropic, and OpenAI, with Irregular noting they are enhancing their methods for safely conducting AI cybersecurity tests. Following Anthropic's disclosure of a similar incident, OpenAI had previously acknowledged that its models gained inappropriate internet access during testing.