PROMPT INJECTION AND TOOL POISONING

CHECK YOUR UNDERSTANDING • Why are language models vulnerable to injection? ◦ Answer: Instructions and data arrive as one stream of text with no guaranteed separation. • Why is indirect injection often more dangerous than direct injection? ◦ Answer: The user never sees it, and agents routinely read content from sources nobody controls. • What is tool poisoning? ◦ Answer: Malicious instructions hidden in a tool's name, description, or metadata. • Name two conditions of the lethal trifecta. ◦ Answer: Access to private data, exposure to untrusted content, and the ability to communicate outward. • Why design as if some injections will succeed? ◦ Answer: Because no method fully prevents them, so limiting what a fooled agent can do limits the harm.