CrowdStrike Tests Model-Weight Scans Against LoRA Backdoors
Lab tests flagged eight poisoned model artifacts; checking a merged model against its base and confirming triggered behavior separates suspicious changes from benign tuning.

CrowdStrike researchers tested a way to spot backdoors in LoRA-tuned models before those models reach an agent with tools. Their November 16 research flagged eight backdoored artifacts in a controlled 16-artifact test, then confirmed the trigger and unwanted behavior in each of the eight. Teams loading third-party open-weight models or adapters into coding assistants and tool-using agents have a new intake question: which base model produced these weights, and what behavior appears when the altered weights are exercised?
One lab model produced an unsolicited tool request toward a canary destination when its prompt contained a trigger. Another generated a look-alike package name, pandas2, instead of the expected library name in example code. Those are controlled demonstrations of conditional behavior, with real consequences depending on what tools a deployed agent can call and what code a user accepts. CrowdStrike writes that a poisoned model “passes standard benchmarks until the trigger appears.”
An adapter hidden inside the base
Hugging Face’s PEFT documentation describes LoRA as a small trainable update layered over a frozen base model. Operators can load that adapter separately or merge it into the base weights to serve a standalone model. A normal prompt may get a normal answer; a chosen word, phrase or date can make a poisoned update steer the model into an attacker-favored output. The merged file can arrive without a separate adapter to inspect.
CrowdStrike’s first stage identifies the base checkpoint and subtracts its weights from the suspect model. It examines where the changes sit across layers, whether their singular values have the sharp rank drop associated with a merged LoRA, and which output tokens the changed directions promote. A declared base that mismatches the weights is itself worth investigating. CrowdStrike can only make a precise weight comparison if the true base is in its clean reference library; hold models with an unmatched base for isolated behavioral testing rather than treating a partial scan as clearance. The resulting score ranks work for an analyst; the scanner’s classification decides which models move to confirmation.
Why a weight flag needs a test
In the lab, all eight backdoored artifacts moved to the second stage: five classified as backdoor-suspect, two as merged-lora, and one as obfuscated-merge. Two benign merged LoRAs also moved forward. One benign merge scored 20 while a backdoored artifact scored 19, so a score alone would have led the analyst astray. CrowdStrike tested its own method on eight lab-built backdoored artifacts, four public controls and four benign lab artifacts.
For an escalated model, the researchers searched candidate trigger inputs by comparing next-token probabilities for the suspect output against a clean alternative. They then generated full responses for promising candidates and checked whether the unwanted behavior appeared. The method found triggers for all eight backdoored artifacts. For the two benign merges, it returned “no trigger found.” CrowdStrike’s search covered four payload types and one prompt shape, so that result is not a general clearance for a different model. A benign merge can carry a strong LoRA signature; the observed behavior is what makes the difference.
The lab also found a caution in refusal testing: some style words weakened refusals in the clean base, and one unrelated public adapter widened that weakness. Test against the same base under the same prompts before assigning every concerning output to a planted trigger.
Make model intake an access decision
Start with the models that can dispatch tools, reach the network or write code someone will run. Record each artifact’s provenance, its exact base checkpoint and whether the deployed file is a standalone adapter, a merged checkpoint or a converted format such as GGUF. Compare a candidate’s weights with the correct base before approving it; a mismatch or low-rank change is a reason to hold the artifact for behavioral testing, especially when the model’s declared origin differs from the base its weights match.
Test flagged candidates in isolation with the same tool declarations and coding tasks planned for deployment. Compare outputs with the clean base, inspect generated package names and tool destinations, and confirm any recovered trigger across several tasks before treating the artifact as backdoored. Keep tool calls sandboxed and monitor agent egress after deployment: those controls limit what a conditional output can reach while the model remains in service. For a newly approved model, the useful verification record pairs the artifact and base checkpoint with the weight-difference result, any triggered-output test and the agent’s tool-access setting. A scanner finding queues a model for testing; the test shows whether its output makes a tool call or changes code under the planned permissions.