Researchers demonstrate self-propagating backdoor in LLMs

phrack.org19 min read

Researchers demonstrate a reproducible "AI Virus" attack where a backdoored large language model used upstream as a training-data helper self-propagates its malicious behavior into independently fine-tuned downstream student models. The implant uses two triggers: one that replicates the malicious training-row structure during data formatting, and another that executes a shell payload when prompted with specific attacker-controlled domains. The findings show that source data and student input prompts remain entirely clean while the compromise is transmitted purely through the helper model's response distribution. Open article →

More on this topic