A Gate Between Reading and Acting

An agent that reads untrusted content and then acts on it needs a check in between, or a hostile input becomes a hostile action.

Gate on the way in and the way out

Classify content before it can steer a tool call, and again before output leaves. The inbound gate protects actions; the outbound gate protects readers.

def decide(verdict):
    if verdict == "unsafe":
        return "block"
    if verdict == "unclear":
        return "flag"
    return "allow"

Record the decision

Whenever a gate stops work, record the decision and the reason so a reviewer can see why the agent halted. For an ambiguous input, abstain or escalate rather than guessing in either direction.