DECEPTION CHECK RESEARCH / AI TOOL ACCESS
Someone is checking for MCP.
If someone sends a request to your MCP service, can your team tell whether it reached a tool? Here’s how to follow one request from your gateway to the tool’s audit log.
The Model Context Protocol (MCP) connects AI applications to tools that can look up data or change records.
Pick one harmless tool in an approved test environment. Run one call you expect the service to allow and one you expect it to deny. Use a separate trusted request ID for each. Find the request in your gateway and MCP server logs, check the access decision, and look for the call in the tool audit. Use the MCP log check worksheet to record what you find.
What the decoys recorded
In a 30-day sample from our research decoys, we verified eight distinct initialize requests sent to POST /mcp. We checked the retained bodies, which used JSON-RPC, MCP’s message format. In pre-July 2026 protocol revisions, initialize was the opening message. An MCP method names the requested action, such as setting up a session or calling a tool.
For most records in the sample, the export found an MCP-like URL and no MCP method in the captured text.
SELECTED DECOY EXPORT · DISTINCT RECORDS
The eight setup requests account for most of the nine records with method text. The remaining record had a bare call label on a different web route and no matching body in the retained evidence. The MCP method for a tool invocation is tools/call.
Four of the verified requests offered protocol version 2025-03-26, and four offered 2025-06-18. We omit the self-reported client names because a requester could imitate those clients and the names could suggest the named software sent the requests. Verify the caller’s identity separately when you review your own logs.
Five of the eight initialization records had HTTP 200 status and three had 404. The retained records include HTTP status but no response bodies. They establish what the clients sent and leave handshake completion and tool use unverified.
Start at the edge. Finish at the tool.
These checks apply to remote MCP services you operate. Use your real routes and service inventory, including proxies and gateways. For local MCP over stdio, start with client and server application logs. There may be no HTTP record.
- 01 / EDGE ACCESS LOG
Did the request reach the right service?
Check the service, route, verb, upstream destination, HTTP status, and UTC time. Use a trusted request or trace ID to match the edge event to the server. Use forwarded source addresses only from proxies you trust.
JSON-RPC request POSTs using July 2026 Streamable HTTP require
MCP-Protocol-VersionandMcp-Method. Three core methods also requireMcp-Name. It carries the name fortools/callandprompts/get, or the URI forresources/read. A configured gateway can record these client-supplied headers. - 02 / MCP SERVER LOG
What operation did the server see?
Check the parsed JSON-RPC method and protocol version. When the server processes the body, the July revision requires it to reject a header and body mismatch. Record the server’s parsed method and validation result.
For an earlier protocol revision, match the successful
initializeresponse to the request and look for the laternotifications/initializedin the same server-side session or connection context. In MCP’s July 2026 revision, clients can send an operation directly, with optionalserver/discover. - 03 / AUTHORIZATION LOG
Was the operation allowed?
tools/listasks which tools are available.tools/callasks to invoke a named tool. Join the method, tool name when present, authenticated principal, authorization decision, and time. For intentionally private services, investigate unexplained allows and establish a baseline for expected tests and denied probes. - 04 / TOOL AND DOWNSTREAM AUDIT
Did the tool do anything?
Check invocation status and the downstream record for the action in question, such as a database read or an object update. Carry a trusted correlation ID into the tool and record failures or partial effects. Mark the outcome unknown when the needed audit trail is absent.
Decide which headers your gateway should retain.
Use a header allowlist. A tool schema can mark arguments with x-mcp-header, which makes their values appear in Mcp-Param-* on conforming calls. Keep that header family out of routine logs.
A server that processes the body decodes Base64-wrapped Mcp-Name and Mcp-Param-* values before comparing them with the request body. If your gateway matches on Mcp-Name, decode wrapped values before matching there too. Log only approved tool identifiers and keep raw resource URIs, which may contain sensitive paths, out of the logs in either form.
Log method, decision, outcome, and references. Keep credentials, session secrets, full tool arguments, and result bodies out of routine security logs.
Example: one request ID across three log sources
Illustrative test records. Field names and outcomes vary by service.
Edge log request=demo-42 route=/mcp http_status=200 Mcp-Method=tools/call Mcp-Name=test_lookup MCP server log request=demo-42 method=tools/call decision=allow Tool audit request=demo-42 tool=test_lookup result=error
The edge recorded HTTP 200 and authorization allowed the call. The tool recorded an error. If the call was expected to read or change something, check the downstream audit for that action before you close the investigation.
Run a safe check before the next alert.
With the service owner, also request the endpoint URL and run an authorized tools/list in the test environment using an approved client. If your service supports initialization from an earlier protocol revision, capture that exchange too. Record the request ID and expected log entries for each test.
When checking the denial, confirm that the authorization decision was logged and that the tool audit covers the test window. If any expected record is missing, record the gap and assign an owner. When a safe test action should produce a downstream record, verify that record independently.
Explore more Deception Check research on what internet-facing systems record.
Sources and methods
The export contained 902 rows for September 8 through October 7, 2026 UTC. Collapsing rows identical across all exported fields removed 51 copies. The category “MCP-like URLs with no MCP method found” covers records with an mcp or sse path segment and no extracted method text in the export. The export searched available request bodies for a method using a text pattern. We checked the eight retained initialization bodies by parsing JSON and matching them to the exported records.
- MCP specification, legacy initialization lifecycle
- MCP project, July 2026 protocol revision announcement
- MCP Streamable HTTP, request headers and validation
- MCP Streamable HTTP, header value encoding
- MCP Streamable HTTP, mirroring tool arguments into headers
- MCP specification, listing and calling tools
- MCP security best practices
- OWASP logging cheat sheet
