# Hidden in the Weights: CrowdStrike scans LoRA merges for backdoors

> CrowdStrike reports that static weight analysis plus trigger recovery found a working trigger in every one of the 8 poisoned LoRA models its lab built.

- **Severity:** Info: Research, tools, evals or lab claims with no direct exposure. Context, not an alert.
- **Category:** Research, Supply chain
- **Published by:** CrowdStrike
- **Disclosed:** 2026-11-16
- **Affects:** Open-weight models, LoRA adapters, Qwen2.5-3B-Instruct, Qwen3-1.7B, Llama-3.2-1B-Instruct
- **Research by:** Donato Onofri (CrowdStrike Cyber Superintelligence Lab), Paul Urian (CrowdStrike Cyber Superintelligence Lab), Jonathan Bar Or (CrowdStrike Cyber Superintelligence Lab)
- **Primary source:** https://www.crowdstrike.com/en-us/blog/supply-chain-risk-from-backdoored-lora-models/
- **Page:** https://theinjection.dev/items/crowdstrike-backdoored-lora-detection/

CrowdStrike's Cyber Superintelligence Lab reports a two-stage scan that flagged all 8 backdoored models its team built, by subtracting the base model's weights and then recovering the trigger from next-token probabilities.

## What happened

CrowdStrike published "Supply Chain Risk from Backdoored LoRA Models" on November 16, 2026. Donato Onofri, Paul Urian and Jonathan Bar Or of its Cyber Superintelligence Lab built 8 backdoored models and 4 benign ones, added 4 public control models, and scanned all 16 artifacts before deployment. The work follows a ProjectDiscovery demonstration from October 2026, which CrowdStrike cites, of a LoRA fine-tune of Qwen2.5-7B-Instruct that made a tool call to an attacker URL when a trigger phrase appeared.

## How it works

Detection runs in two stages and does not need the trigger in advance. Stage one identifies the base model, subtracts its weights from the suspect model and reads the difference three ways: where updates concentrate across layers and modules, whether the update carries a LoRA-style rank ceiling, and which output tokens the update promotes. Stage two takes flagged models and uses only next-token probabilities, stopping generation at the point where the backdoor makes its choice, such as which library to import, and testing whether a candidate input shifts probability sharply toward the malicious token. Top candidates are confirmed with full generations and trimmed to the shortest working trigger, or the tool reports no trigger found.

## Who is affected

Anyone importing open-weight models, merged checkpoints or LoRA adapters is in scope, because a merged adapter is not visible from a model card and a cheap fine-tune can carry a dormant payload. The authors group the risk into agent backdoors that run commands or call attacker infrastructure, code backdoors that emit malicious or subtly exploitable code, and safety backdoors that drop refusals on a trigger. They state the limits of their own results: 16 models tested, so thresholds and false-positive rates for public community merges are unknown, and full fine-tunes and direct weight edits were not addressed.

## What to do

Scan every model statically as it enters your pipeline, and escalate undisclosed merged LoRAs, unusual layer profiles and payload-shaped vocabulary to trigger recovery before you deploy. CrowdStrike says to block, quarantine or approve on the evidence rather than on the model card, and to keep runtime controls such as sandboxed tool calls and egress monitoring, which act after deployment. Test refusal behaviour against a clean base model too: the team found that some style words relaxed refusals even on the clean base, and that two unrelated public adapters widened that weakness.

## Quick facts

- **Published:** November 16, 2026, by CrowdStrike's Cyber Superintelligence Lab
- **Scanned:** 16 artifacts: 8 backdoored, 4 benign in-house, 4 public controls
- **Stage 1 result:** All 8 backdoored models flagged, 5 of 8 benign reported clean
- **Stage 2 result:** A trigger confirmed for all 8 backdoored models, no trigger found for both escalated benign merges
- **Base models:** Qwen2.5-3B-Instruct, Qwen2.5-Coder-1.5B-Instruct, Qwen3-1.7B, Llama-3.2-1B-Instruct
- **Formats:** safetensors, GGUF and LoRA adapters, merged and unmerged
- **Stated limits:** Full fine-tunes and direct weight edits not covered

## Sources

1. [CrowdStrike Research](https://www.crowdstrike.com/en-us/blog/supply-chain-risk-from-backdoored-lora-models/)
2. [ProjectDiscovery Demo](https://projectdiscovery.io/research/how-abliterated-models-can-get-you-pwned)
