FlashRT: An Efficient Red-Teaming Framework for Prompt Injection and Knowledge Poisoning in Long-Context LLMs

The researchers propose the FlashRT framework to improve the efficiency of optimization-based prompt injection and knowledge poisoning attacks on long-context LLMs while reducing computation and GPU memory overhead. According to the paper, compared with the SOTA baseline nanoGCG, FlashRT achieves 2x-7x speedup (for example, runtime reduced from one hour to under ten minutes) and 2x-4x GPU memory reduction (for example, 32K token context reduced from 264.1 GB to 65.7 GB). The framework can be broadly applied to black-box optimization methods such as TAP and AutoDAN; the authors hope it can serve as a red-teaming tool to support systematic evaluation of long-context LLM safety. Code has been released.