Skills
Unit 42: 22 Distinct Indirect Prompt Injection Techniques Identified from Real-World Telemetry Analysis
Palo Alto Unit 42 analyzed real-world production telemetry to identify and systematize 22 distinct techniques of indirect prompt injection used against deployed AI agents. Their formal decomposition breaks LLM agent security into four core properties and demonstrates why pattern-matching defenses fail and how current benchmarks miss key attack vectors. This is the first large-scale empirical taxonomy of prompt injection attacks from production data rather than lab conditions, providing defenders with a concrete threat model to test against.
Source
↳ Follow the thread