Cybersecurity researchers from Tenable have identified critical vulnerabilities in Google’s Gemini AI platform, collectively known as the “Gemini Trifecta.” These flaws expose users to prompt injection attacks and data leakage across key features, including Cloud Assist, Search Personalization, and Browsing Tool. The first vulnerability enables attackers to manipulate Google Cloud logs, executing unauthorized commands through injected instructions. The second flaw allows attackers to exploit users’ Chrome search histories, effectively turning it into a malicious prompt that can exfiltrate sensitive information. Lastly, the Browsing Tool vulnerability lets attackers capture sensitive user data in requests to their servers without detection. Google has patched these issues following responsible disclosure. This research underscores the need for rigorous input validation and access restrictions in AI systems. Organizations utilizing similar technologies are encouraged to monitor model behavior and limit external tool access to mitigate risks associated with user data processing.
Source link
