Three Methods Found to Protect LLMs from Backdoors

Researchers at NC State University (Jianwei Li, Jung-Eun Kim, Min-Seon Kim) developed three methods to protect LLMs from backdoor attacks, which, according to the authors, reduce the attack success rate from approximately 100% to 0–10% across two tasks, three attacks, and four model families.

Compiled by Sergey KostenchukPublished 2026-11-08Updated 2026-11-08

2026-11-08 Safety

🛡 Three Methods Found to Protect LLMs from Backdoors

The NC State University team built protection at three levels of the model: null-space projection filters LoRA updates, QES isolates trigger behavior in a separate expert, and the immunization approach seeks a common “backdoor signature.”

🌍 Fine-tuning in finance and software development is growing, and a backdoor in a LoRA adapter remains undetected until an attack. The methods clean the model without knowledge of the trigger, references, or retraining; according to the authors, ASR drops from ~100% to <10% (null-space) and 0–10% (QES).

👤 A fine-tuned model can undergo a pre-release check: a drop in ASR to near zero is a concrete benchmark, and the Secret Alignment concept proposes triggers as legal authorization.

Source 1: https://news.ncsu.edu/2026/11/new-techniques-to-prevent-backdoor-attacks/ Source 2: https://www.nationaltribune.com.au/researchers-demonstrate-comprehensive-suite-of-new-techniques-to-protect-ai-from-backdoor-attacks/

Sources