Agent

Agents are LLM applications that can take actions — call tools, write files, hit APIs, browse the web, or invoke other agents. That capability shifts the security model fundamentally: a prompt-injection payload in a chat app is annoying, but the same payload in an agent can trigger real actions (send email, transfer funds, push code). Indirect prompt injection is especially dangerous here because agents routinely consume untrusted content (web pages, emails, files) where attacker instructions can hide. The OWASP LLM Top 10 added "Excessive Agency" as LLM08 specifically for this class. AI Threat Alert tracks CVEs in popular agent frameworks (LangGraph, CrewAI, AutoGen, AutoGPT, LangChain agents) and incident reports from AIID for production agent misuse. Defenses: human-in-the-loop for irreversible actions, scoped tool permissions, separate trust boundaries between agent-controlled and user-controlled context, and budget caps on tool invocation.