1-minute read
What a defender needs to know
CrowdStrike tested weight inspection against conditional backdoors in LoRA-tuned open-weight models. Its scanner escalated all eight lab-built backdoored artifacts in a 16-artifact test. A second stage recovered triggers and confirmed the unwanted behavior in all eight. Two benign merged LoRAs also reached that second stage, where testing returned “no trigger found.” Use a weight anomaly to queue a behavioral test of the model.
LoRA adds a small trainable update to a frozen base model. An adapter can be distributed separately or merged into the base weights, leaving one standalone artifact. CrowdStrike compared each suspect model against its base to locate changes and inspected whether those changes pointed toward concerning output. In one controlled example, a triggered model produced an unsolicited agent tool request; another steered generated code toward the look-alike package name pandas2.
For models that can call tools or write code, record the deployed artifact and specific base checkpoint, then compare their weights. Hold mismatched or suspicious artifacts for isolated behavioral tests using the intended agent tool declarations. If the true base is missing from the clean reference library, the weights cannot be precisely diffed; keep that model on the test path. Where the clean base is available, compare outputs; inspect destinations and package names, and confirm the behavior across tasks. Keep tool calls sandboxed and monitor egress in deployment. Record the artifact, base, weight result, triggered-output result and tool permissions together so the approval applies to the model actually put to work.