Securing agents that can actually take action

An agent that can only talk to you is limited in how much damage a bad response can do. An agent that can call an API, send an email, or execute a transaction is a different risk profile entirely.

The same detection principles apply — score the response before it's acted on — but the stakes and the latency budget both change. You need a verdict fast enough to sit inline before a tool call fires, not just a report generated after the fact.

We walk through how we adapted response-level scanning for tool-calling agents, including what we chose to check synchronously versus asynchronously.