hacnho/tensorrt-detectionlayer-nan-score-bypass-poc

TensorRT DetectionLayer NaN score_threshold bypass proof of concept

This repository contains a bounded research PoC for TensorRT (.engine / .trt / .mytrtfile).

The security question is whether a serialized DetectionLayer_TRT payload can bypass the creator-side score_threshold >= 0 check by carrying NaN, while the final built engine still loads and executes normally and silently suppresses detections.

Files

• control.engine
• nan-score.engine
• verify_remote_poc.py

What the files demonstrate

Control:

positive_cls -> [0.5]
mixed_bbox   -> [0.11951626092195511]

Malicious serialized score_threshold = NaN:

positive_cls -> [0.0]
mixed_bbox   -> [0.0]

Verify the public HF artifacts

After unauthenticated download, run:

python verify_remote_poc.py

Expected result:

• both engines deserialize successfully
• both engines execute successfully
• the nan-score.engine output differs from the control on normal presets
• semantic_suppression_observed is true