Cyber Risks · Trail of Bits Blog

Assessing the security posture of a widely used vision model: YOLOv7

Summary: Trail of Bits identified 11 security vulnerabilities in YOLOv7, the open-source object detection codebase from Academia Sinica with over 3,000 forks, including five high-severity and three medium-severity findings that could enable remote code execution, denial of service, and model differentials. The most notable finding is insecure parsing of YAML model configuration files, where the parse_model function calls eval on unvalidated file contents, and the maintainers did not respond to disclosure attempts, leaving the issues unfixed at publication.

Key Takeaway: YOLO is the de facto algorithm in object detection and is used in robotics, autonomous vehicles, and manufacturing, so a codebase its own reviewers call unsuitable for mission-critical use is already running in systems where a manipulated YAML file can execute attacker code.

Read the original at Trail of Bits Blog ↗

Topics
Endpoint & Network Security
Related