ModelScan logging.config.dictConfig bypass PoC
Private security disclosure artifact for the Huntr Model File Vulnerability program.
Contents
VULN_REPORT_2.md- technical vulnerability reportpocs/create_logging_dictconfig_pickle.py- PoC model-file generatorpocs/load_logging_dictconfig_pickle.py- local loader that demonstrates executionpocs/logging_dictconfig_bypass.pkl- malicious pickle model filepocs/logging_dictconfig_bypass.joblib- malicious joblib model filepocs/modelscan_logging_dictconfig_pkl.json- local ModelScan result for the pickle filepocs/modelscan_logging_dictconfig_joblib.json- local ModelScan result for the joblib file
Safety
These files are intentionally malicious proof-of-concept model files. Do not load the .pkl or .joblib files outside an isolated test environment.
The PoC was locally verified against ProtectAI ModelScan tag v0.8.8 at commit 61fcec9c2a37c24c1fb12d84ede30fe248a364bd: ModelScan reported zero issues for both files, and loading the files with Python 3.12 executed the embedded benign marker command.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support