LLM agent security needs more than guardrails

Madhu Shashanka, Chief Scientist at Concentric AI, explains why securing LLM tools and agents depends more on your data than on the model. Controls on model behavior, such as guardrails and injection detection, lower the rate of bad outcomes but never reach zero. Controls on what goes in are different, because you own the documents, systems, and credentials an agent can reach.

He covers prompt injection, data leakage through connectors, overprivileged agents, and shadow AI. For each risk, he describes what the model side control does and where it stops. A file an agent cannot read cannot leak, he says, so access decisions set the limit.

He then sets out three layers: input controls, behavioral controls, and output and action controls. Topics include sensitivity classification, fixing oversharing before connecting a copilot, giving agents their own identities, logging every tool call, and requiring human approval for actions that cannot be undone.

Download: The IT and security field guide to AI adoption