Input Filtering

LLM prompt injection defenses are bypassable, even with advanced techniques

Prompt injection attacks exploit the fundamental nature of LLMs where instructions and data are indistinguishable within the context window. While various defense layers exist, from simple keyword filtering to using a s…