On 16 November 2026, CrowdStrike described a two-stage method for finding backdoors in open-weight models modified with low-rank adaptation (LoRA). The company says the method first examines model weights for signs of modification, then searches for inputs that trigger suspect behavior.[1]

LoRA fine-tunes a model by training a smaller set of additional parameters, known as an adapter. The adapter can be distributed separately or merged into the original weights. CrowdStrike says a backdoor can leave a model behaving normally until a particular input activates unwanted behavior, such as a malicious tool call or insecure code.[1]

The first stage compares a suspect artifact with its base model and examines the differences in shape, rank and promoted output tokens. CrowdStrike reports that it scanned 16 artifacts in a controlled lab: eight backdoored models, four benign models built in the lab and four public controls. It says all eight backdoored models were escalated for further investigation. Two benign merged models were also escalated, while another control was classified as an abliterated model. The company notes that a malicious merged LoRA and a benign one can look alike to the scanner.[1]

The second stage probes escalated models for candidate triggers, using next-token probabilities and checking candidates with full generations. The supplied source describes this procedure, but its saved text ends as the stage-two results begin, so those results cannot be established from the material available here.[1]

Sources

  1. Hidden in the Weights: Supply Chain Risk from Backdoored LoRA Modelswww.crowdstrike.com · 16 Nov 2026

Check or correct this page

Verification record

The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.

ec7e056c2764b88e40d12af0b76b80903bc49e8dc346ef75e44ee75ebe652742