Yoloe
Monthly
Code execution in THU-MIG YOLOE's AutoBackend model loader allows a local attacker to run arbitrary Python by supplying a crafted model whose imgsz, names, or kpt_shape metadata is passed to eval() in ultralytics/nn/autobackend.py; the victim must load the untrusted artifact, so this is not a network-facing trigger. Publicly available exploit code exists, but there is no CISA KEV confirmation, and the combination of local access, crafted-file prerequisite, user interaction, and limited deployment of this research project yields LOW real-world priority.
Code execution in THU-MIG YOLOE's AutoBackend model loader allows a local attacker to run arbitrary Python by supplying a crafted model whose imgsz, names, or kpt_shape metadata is passed to eval() in ultralytics/nn/autobackend.py; the victim must load the untrusted artifact, so this is not a network-facing trigger. Publicly available exploit code exists, but there is no CISA KEV confirmation, and the combination of local access, crafted-file prerequisite, user interaction, and limited deployment of this research project yields LOW real-world priority.