Rammadaeus/mmdetection-ace-poc

PoC: mmdetection Multiple Arbitrary Code Execution Vulnerabilities

Bounty: https://huntr.com/bounties/ca85b6b6-b866-4f79-bc52-68f8daf7b0d1
Severity: Critical (CVSS 9.8)
Affected: open-mmlab/mmdetection (all versions, 29k+ stars)
Reported: 2026-02-17

Vulnerabilities

#	Severity	File	Line	Issue
1	CRITICAL	mmdet/models/task_modules/tracking/camera_motion_compensation.py	26	eval(warp_mode) on config param
2	CRITICAL	tools/analysis_tools/analyze_logs.py	207	eval(args.task) on CLI arg
3	CRITICAL	tools/analysis_tools/benchmark.py	128	eval(f'{args.task}_benchmark')
4	HIGH	tools/model_converters/*.py (7 files)	various	torch.load() without weights_only
5	HIGH	mmdet/datasets/openimages.py	483	np.load(allow_pickle=True)
6	HIGH	mmdet/evaluation/metrics/coco_occluded_metric.py	69-70	Remote .pkl URL deserialization

Reproduction

pip install mmdet mmengine opencv-python torch numpy

# Vector 1: eval() on config-controlled parameter (most critical)
python poc_eval_config.py

# Vector 2: torch.load() without weights_only
python poc_torch_load.py

# Vector 3: np.load(allow_pickle=True)
python poc_numpy_pickle.py

Root Cause

The most impactful finding is eval(warp_mode) in CameraMotionCompensation.__init__(). The warp_mode parameter flows from mmdetection config files, which are the standard sharing artifact in the ecosystem (model zoos, academic papers, tutorials). A single malicious config achieves RCE at model init time with no authentication or user interaction beyond loading the config.

Fix

Replace eval() with an explicit dispatch dict:

_WARP_MODES = {
    'cv2.MOTION_TRANSLATION': cv2.MOTION_TRANSLATION,
    'cv2.MOTION_EUCLIDEAN': cv2.MOTION_EUCLIDEAN,
    'cv2.MOTION_AFFINE': cv2.MOTION_AFFINE,
    'cv2.MOTION_HOMOGRAPHY': cv2.MOTION_HOMOGRAPHY,
}
self.warp_mode = _WARP_MODES[warp_mode]  # KeyError on invalid input