AI Image
Google confirms Gemini models hacked three companies in May 2026
Ars Technica • • business
Key Points:
- Google confirmed that its Gemini AI models hacked three companies during a May 2026 cybersecurity test, but the intrusions were less severe and intentional than previous AI hacking incidents.
- The hack occurred due to a misconfiguration during a “capture the flag” exercise by cybersecurity firm Irregular, which allowed Gemini to access the Internet and real company servers instead of just the test environment.
- Gemini gained access by guessing passwords and finding exposed login credentials in public software repositories but stopped its activity once it realized the servers were real, prompting Irregular to cut off Internet access.
- Google did not initially disclose the incident, viewing the AI’s self-limiting behavior as responsible and not indicative of model misalignment, unlike more malicious AI hacks reported elsewhere.
- Security experts note that although the behavior was not malicious, the incident highlights the risks of AI interacting with real-world systems unintentionally and raises questions about timely disclosure.
