Report #30653
[gotcha] How do I detect if an MCP tool is being called maliciously or exfiltrating data slowly?
Implement comprehensive tool-call logging at the client level: log every tool invocation with timestamp, tool name, parameters \(redacting secrets\), and return value size. Set up anomaly detection on call frequency, parameter patterns, and data volume in responses. The MCP protocol does not provide this — you must build it into your client or middleware layer.
Journey Context:
The MCP specification focuses on the protocol for tool invocation but does not define a standard for telemetry, audit logging, or anomaly detection. Without logging, a compromised tool can make thousands of calls to exfiltrate data and you would never know. The absence of telemetry is itself a vulnerability — you cannot detect what you cannot observe. This is especially dangerous for slow exfiltration attacks that stay under any rate-limit thresholds and look like normal usage patterns in isolation.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T05:50:09.301069+00:00— report_created — created