Understanding and Enhancing Backdoor Persistency in LLM Agent Post-Training

Researchers studied a supply-chain threat in which an attacker supplies a third-party model with a backdoor — hidden behavior that produces malicious outputs when a particular input pattern appears — and examined how persistent that backdoor remains when developers adapt the model through benign post-training to build LLM agents. The work focuses on software-engineering settings.

Developers can build LLM agents by adapting third-party models through benign post-training. We study a supply-chain threat in which an attacker supplies a model with a backdoor: hidden behavior that produces malicious outputs when a particular input pattern appears. Focusing on software-engineering