AI Agents Expose Sensitive Data in Half of Security Tests

Recent security assessments of AI agents have uncovered substantial vulnerabilities, with one AI model reportedly divulging sensitive information in roughly 50% of attempted breaches. This development adds to growing concerns surrounding AI's involvement in security incidents, including accusations of its use in bank hacks and data breaches.

Researchers identified a novel prompt-injection method that encrypts malicious commands, allowing them to circumvent initial security filters. The AI subsequently decrypts these instructions, potentially granting unauthorized access to local secrets and facilitating the exfiltration of data. The inconsistent success rates observed across various AI models, including those used by Copilot, cast doubt on the efficacy of AI as a robust security barrier, emphasizing the critical need for stringent controls around AI functionalities rather than solely relying on the models themselves.