Research
Understanding and Enhancing Backdoor Persistency in LLM Agent Post-Training
Publication date unknown · Discovered arxiv.org
Page published
Publication date unknown · First observed: 17 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
PERSISTBD shows that developers adapting third-party models into agents cannot rely on benign post-training alone to eliminate inherited backdoors.
The PERSISTBD study demonstrates that backdoors in software-engineering agent models can survive benign supervised fine-tuning (SFT) and subsequent reinforcement learning (RL). On Qwen2.5-Coder-7B, the proposed refinement method increased attack success from 20% to 74% after SFT and from 20% to 76% after SFT–RL, while maintaining comparable benign task performance.