Indirect prompt injection through retrieved content
How instructions embedded in documents, email and web content can redirect LLM-based assistants and agents, with layered mitigations across input handling, tool permissions and output validation.
Read the advisory




