Vulnerability / research

CrowdStrike detects hidden behavior in eight deliberately altered model artifacts

Reported

Published here

What happened

CrowdStrike tested whether defenders could distinguish tampered open-weight models and adapters from controls before deployment. The researchers combined inspection of model weights with behavior tests in an isolated environment.

Parties and reporting

Where this happened

Research, tests and decoys. Lab studies, demonstrations, sandboxed evaluations, honeypots and disclosed flaws without a documented real-world victim. There is no victim location to map. Explore the world map.

AI tools and systems: Low-rank adaptation (LoRA) models and adapters, Qwen and Llama model families

Outcome

The team confirmed a trigger in all eight backdoored artifacts, which represented seven distinct backdoors. Follow-up tests cleared the two benign artifacts flagged in the first stage.

Operating lesson

Compare downloaded weights with a trusted base model and test suspicious behavior in isolation before deployment.

Evidence and limits

All eight backdoored artifacts were prepared for the experiment. The 16-artifact sample does not measure how often backdoors occur in public models or establish compromise of a production deployment.

Sources