CrowdStrike describes backdoored LoRA fine-tunes risks

crowdstrike.com3 min read

CrowdStrike researchers outline the supply chain risks associated with backdoored open-weight models, focusing on Low-Rank Adaptation fine-tuning as a low-cost attack vector. The authors explain that adversaries can plant stealthy backdoors—such as agent, code, or safety triggers—into models that otherwise pass standard evaluations. To address this threat, the research team developed a two-stage approach combining static weight analysis and trigger recovery to detect and confirm backdoors before deployment. Open article →