Research
AutoBackdoor: Automating Backdoor Attacks via LLM Agents
Publication date unknown · Discovered aclanthology.org
Page published
Publication date unknown · First observed: 13 Nov 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
AutoBackdoor demonstrates how agents can automate training-data poisoning, highlighting the need to validate fine-tuning data and test customized LLMs for hidden, trigger-dependent behavior.
AutoBackdoor is a research framework that uses LLM agents to generate backdoor triggers, construct and refine poisoned training samples, and produce backdoored models through fine-tuning. The paper reports attack success above 90% using a small number of poisoned samples in experiments covering LLaMA-3, Mistral, Qwen, and GPT-4o, with limited protection from the defenses evaluated.