Backdoor attacks against Machine Learning IRL
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
In Phrack, G. Cherubin describes a parser flaw in ONNX Runtime: sparse tensors are densified at load time, before external-path validation, so a model file can make the runtime read arbitrary files or memory addresses. Combined with graph logic in a Qwen2.5-7B-Instruct model, a trigger word makes the model emit the read bytes as tokens, with no change to the runtime. Both fixes shipped in onnxruntime v1.27.0. The author's point for builders: downloaded models are code and need the same vetting as libraries.