How fast does an AI agent guardrail need to be?
Fast enough to rule before the action completes, which a language model on its own usually isn't. In our prototype an LLM ruling on every system call an AI coding agent made timed out on 56% of events at a 4-second budget. Moving clear-cut patterns to a synchronous deny-list in front of the model cut the p99 for those events from 3,617 ms to 1 ms.
Answered from GDF Research experiments by Sterling Morrison. Numbers link to the papers below.
The setup
The prototype intercepts the file, network and process system calls of an agent's process tree on Windows and asks Claude Haiku 4.5 on Amazon Bedrock to allow or block each one. We measured 1,247 events and 1,000 threat scenarios.
What decided latency
Placement, more than the model. Unambiguous high-risk patterns, such as dumping the Windows password hive, fetching a payload with certutil, or opening a raw reverse-shell socket from PowerShell, don't need a model's judgment. Checking them synchronously on the publishing thread, before anything is queued for review, removed them from the slow path entirely and left the model to rule on the cases that actually need judgment.
The caveat
These numbers cover the userspace verdict pipeline with the kernel drivers unloaded on the test host. End-to-end enforcement through the drivers, and an adversarial test against a prompt-injected agent, are still to come.
