Clouding the Mirror: Stealthy Prompt Injection Attacks Targeting LLM-based Phishing Detection
This research evaluated prompt injection (PI) attacks against multimodal LLMs used for phishing detection. The study developed a two-dimensional taxonomy to classify attack techniques and surfaces, testing several models like GPT-5 and Llama 4. A new defense framework, InjectDefuser, combining prompt hardening, allowlist RAG, and output validation significantly reduced attack success rates across all tested models.