The tool is still called publish_report. Yesterday it accepted a report ID and wrote to an internal workspace. Today its schema also accepts a destination. The name is familiar. The authority being requested may be different.
This is a worked design example, not a reported incident or a completed Ackuity test. Its purpose is to make a review question concrete: what exactly did the person approve, and what will the target receive?
DISCOVERY IS NOT APPROVAL
What can change between discovering a tool and calling it?
The versioned MCP tools specification describes tool names, descriptions and input schemas. It also defines a tool-list change notification for servers that declare the corresponding capability. Clients discover the available tools through tools/list. See the MCP tools specification.
Those mechanisms describe the interface. They do not establish that a specific user authorized every operation that interface can express. A description may change, a schema may expand or the implementation behind an unchanged definition may behave differently. These are separate conditions and need separate checks.
Tool metadata can also influence what the model chooses to do. A malicious description can try to redirect the workflow, a form of tool poisoning. A harmless wording correction is different. Treating every change as an attack would create avoidable review noise.
ONE TOOL, TWO CONTRACTS
Why does the same tool name not preserve the same approval?
Imagine a person approves publishing synthetic report R17 to the internal review workspace. The proposed invocation later names an external destination. The original approval tells us who approved the internal publish. It does not answer whether the external publish is allowed.
- Definition: which operation and inputs were available when the request was reviewed?
- Invocation: which values, resource and destination are proposed now?
- Authority: which policy and approval cover that exact operation?
- Outcome: what did the receiving system actually accept or change?
Our recommendation is to bind review to a defined operation and its significant parameters, not just a friendly tool name. This applies OWASP's transaction-authorization principle to an agent workflow. It is Ackuity's design analysis, not a claim that MCP mandates this approval format. OWASP transaction authorization guidance explains why authorization must remain tied to the operation being performed.
A PRACTICAL CHECK
How should you recheck a changed MCP tool?
- Record the reviewed contract. Keep the server identity, tool name, relevant definition, target, parameters and policy version. A definition fingerprint can help detect change; it cannot prove the server's behavior.
- Classify the difference. Separate cosmetic metadata changes from changes to resources, destinations, data access, side effects or delegated authority. The classification needs a trusted owner, not the tool's own claim that it is safe.
- Re-evaluate the proposed call. Compare its actual parameters and target with the current policy and the approved operation. Unknown coverage stays unknown.
- Request review where permitted. A material change can require renewed human review. A hard policy prohibition still blocks the operation; a person clicking approve should not silently override it.
- Check the target. Correlate the decision with an independent receiving-system record. A notification, a refusal message or a blocked verdict alone does not establish the target outcome.
MCP's advice that tool annotations should be treated as untrusted unless from a trusted server reinforces the distinction between a hint and an enforcement decision. The MCP security guidance also addresses implementation risks such as confused deputies and token handling. Check the specification version supported by your client and server.
KEEP USEFUL WORK IN THE TEST
What should the evaluation worksheet test?
Use an isolated workspace, synthetic data and a test identity. Write expected outcomes before running anything. These four cases are proposed tests, not measured results:
- Unchanged permitted call: the approved internal publish should still complete.
- Cosmetic description change: evaluate whether the implementation can avoid an unnecessary block while preserving the operation's scope.
- Changed destination or broader input: the earlier approval must not automatically authorize a new operation. Apply current policy and review requirements.
- Unchanged definition, changed server behavior: check independent target records. A matching fingerprint is not sufficient evidence of safe execution.
Record the request, decision and target result under a common correlation reference. If target evidence is missing, record the outcome as unknown. Include missed blocks and mistaken blocks so the exercise measures both safety and useful work.
Open the MCP tool-change worksheet. Fill it in locally and use your browser's print option to save a PDF. The worksheet sends no answers to Ackuity and does not save your entries after you leave.
WHERE THE LAYERS FIT
Does signing the MCP server solve this problem?
Signing and release verification can help establish provenance and integrity for the artifact checked. They do not settle whether the next invocation is authorized for its data, destination and task. Our signed MCP server analysis covers that data boundary. This article covers a different boundary: what changes after discovery or approval.
Ackuity is the Agent Execution Control Switch. Its Agent Security Context Graph brings action context into supported execution paths. Whether a deployment can inspect a changed tool definition, bind an approval or intercept a particular call must be confirmed for that integration. This article does not demonstrate those capabilities or claim universal MCP coverage.
Use the AI agent security platform evaluation guide to compare actual routes and evidence. For mistaken blocks, pair this exercise with the false-positive evaluation method.
SOURCES AND LIMITS
What supports this analysis?
The protocol descriptions above use the MCP 2025-11-25 tools specification and its security guidance. The operation-bound approval recommendation draws on OWASP transaction authorization. The scenario, change classification and proposed tests are Ackuity analysis. No customer, product benchmark or completed prevention result is asserted.
