FlashRT: Computationally Efficient Red-Teaming for Long-Context LLM Prompt Injection

arXiv·high signal

Addresses a critical gap in LLM security: red-teaming long-context models (Gemini 3.1 Pro, Qwen 3.5) for prompt injection and knowledge corruption is prohibitively expensive with existing methods. FlashRT introduces a memory-efficient approach that reduces compute cost by an order of magnitude while maintaining attack discovery rates. Essential for anyone deploying RAG systems or long-context agents who need to stress-test against adversarial inputs without burning GPU budgets.