Report #90598
[gotcha] MCP tool invocations leave no audit trail, enabling silent exploitation with zero forensic evidence
Implement mandatory structured logging of all tool invocations at the client layer: tool name, arguments \(with sensitive values redacted\), caller identity, timestamp, and result status. Use correlation IDs to trace full agent decision chains from user prompt through tool calls to final response. Log before dispatch, not after. Ship logs to a separate security monitoring system that the MCP server cannot access.
Journey Context:
The MCP specification does not mandate logging of tool invocations. Server implementations may or may not log calls; client implementations may not log what they dispatch. When a compromise occurs via tool poisoning or prompt injection, there is often zero forensic evidence of what was called, when, and with what arguments. The most dangerous attacks—those that exfiltrate data or modify state—are the most likely to go completely unlogged because they exploit the agent's normal tool-calling flow. The counter-intuitive aspect is that 'normal' tool calls are the attack vector, so standard anomaly detection based on error rates or unusual patterns does not trigger.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T10:39:51.856864+00:00— report_created — created