A nearly undetectable LLM attack needs only a handful of poisoned samples

This report summarizes research on ProAttack, a clean-label prompt-based backdoor for NLP models that assigns malicious prompts to target-class training samples so that any input carrying the prompt triggers a chosen output. The attack achieved near-100% success across multiple benchmarks and low-data settings while maintaining clean accuracy; several existing defenses were inconsistent, and the authors propose low-rank parameter-efficient fine-tuning (e.g., LoRA) as an effective but task-sensitive mitigation.