AI agent security

AI agent security is the practice of protecting AI agents from being subverted and protecting everything else from the agents: screening the inputs that steer them, governing the actions they take, containing the blast radius when one is compromised, and keeping evidence of what every agent did.

It differs from classic application security in where the attack surface sits. The agent's "code" includes every piece of text it reads: prompts, tool descriptions, web pages, tickets, files. Any of them can carry instructions. And the agent's failure mode is an action, so the control point that matters most is the moment before an action executes.

The discipline layers several controls: input screening, tool vetting, scope enforcement, session-level monitoring, and audit evidence. The agent security stack maps how they fit together.

Related

• The agent security stack
• The 17 agentic attack techniques
• All glossary terms