Definition
Plain language
A monitored checkpoint every action passes through, which records both what a tool really returned and what the AI was actually shown.
As stated in the literature
An instrumented chokepoint for all agent tool calls that separately logs the true tool output and the (possibly tampered) content presented to the agent, enabling detection of indirect attacks that poison tool results.
Why it matters: Without this comparison you can't tell whether an AI misbehaved on its own or was fed poisoned information by an outside tool.
For example, if a web page secretly contains hidden instructions, the gateway records both the clean page and the tampered version the AI ended up reading.
Heard on the show
“And fourth, the tool gateway — a monitored chokepoint that every action flows through, which records two separate things: what the tool really returned, and what the agent was actually shown.”Episode 202 — How Do You Know an AI Agent Actually Refused? Check the World, Not the Words