Study finds privilege-confusion attacks feasible on six of seven AI accelerators

Researchers reported that malicious applications could trick six of seven tested edge AI accelerators into performing privileged operations, potentially affecting more than 128 system-on-chips and 100 million devices. The preprint examined hardware from Google, Nvidia, Hailo, Texas Instruments, NXP, AWS and Rockchip using DeputyHunt, a framework that combines static and dynamic analysis with a large language model. The vendors acknowledged the findings, which received CVE-2025-66425. A proposed on-demand validation defense produced roughly 15% average runtime overhead in a Gem5-salam simulation, compared with roughly 56% for traditional defenses; the evaluation was simulated rather than a production deployment.