Tool-Definition Integrity
A tool's definition — its description, its input and output schema, its annotations — is instruction the agent reads and acts on. An MCP server can change that definition at any time, after you have reviewed and approved the tool. A tool that was safe when you granted it can quietly become one that tells the agent to do something else. Tool-definition integrity closes that gap: AgentPass remembers the definition you approved, and quarantines a tool the moment its definition changes, until a person reviews what changed.
What Happens When a Definition Changes
AgentPass keeps the approved definition of every enabled tool. When it next observes that tool, it compares the two:
- The definition is unchanged. The tool continues to work exactly as before.
- The definition has changed. The tool is quarantined. The new definition is kept aside as pending review; the approved copy is what the agent continues to see until an administrator decides. A tool-definition-changed event is recorded so the change is visible in your audit trail and in the guardrail surfaces.
While a tool is quarantined, it is removed from the tool list the agent is shown and refused if called — so a changed tool cannot be used, discovered, or routed to in the window before a person looks at it.
A quarantined tool is unavailable to everyone until it is reviewed — there is no role that bypasses it, and an administrator's session is held to the same rule as any other. This is deliberate: the whole point is that nobody acts on a changed tool before a person has seen the change.
Reviewing a Change
An administrator reviews a quarantined tool from its connector's tools view. The review shows the approved definition and the pending one side by side, so you can see exactly what changed, and offers two outcomes:
- Approve — the pending definition becomes the new approved definition and the tool returns to normal availability (still subject to the role grants that already govern it).
- Reject — the change is refused; the tool stays quarantined.
Either outcome is recorded with who decided and when, so a definition change becomes an audited administrative decision rather than a silent overwrite.
Why It Matters
Three of the attack patterns input guardrails defend against work by changing a tool after approval — a tool that behaves one way when you grant it and another way later, a description edited to carry hidden instructions, or a tool swapped for a different one behind the same name. Pinning the approved definition and quarantining any change turns all three into a review step you control. You can see these patterns demonstrated, blocked, in The Proving Ground.
Related
- Input Guardrails Overview
- Controlling MCP Tool Access by Role — the role grants that still apply after a tool is approved
- Connecting MCP Servers — where connector tools come from
- Security Controls Reference