How Rogue-Agent Defense works
Rogue-Agent Defense is an AI agent firewall — a layer of protection that sits between autonomous AI agents and the damage they can cause when something goes wrong. AI agents are moving from chatbots to actors: they read inboxes, book meetings, run code, and move money. Every one of those capabilities is a surface for attack. This page documents the four defensive layers and how they compose.
Layer 1 — Prompt-injection shield
A single crafted message — a hidden instruction on a webpage the agent reads, a poisoned email in the inbox it triages — can redirect an agent. The shield scores inbound messages before the agent ever sees them; suspicious input is quarantined or blocked, and the score and the reasons travel with the message when it is passed through.
Layer 2 — Policy enforcement
Per-action rules decide which tool calls execute automatically, which need human approval, and which are denied outright. The policy is evaluated at action time, on the concrete call the agent is attempting — not on the agent's stated intentions.
Layer 3 — Canary tripwires
Planted secrets that no legitimate workflow should ever touch. If a canary fires, an agent has gone off-script — the tripwire converts a suspicion into a signal with a timestamp and a culprit action.
Layer 4 — Kill switch
One click halts an agent fleet when something looks wrong. The other layers exist to make the kill switch a rare event; the kill switch exists so that when they miss, a human can stop the damage in one action instead of chasing individual agents.
Honest scope
The layers defend the paths they can see: messages routed through the shield, actions routed through the policy gate, secrets planted as canaries. An agent wired around the layers is outside their protection — the defense is the wiring, kept complete.