Report #1547
[gotcha] Unable to detect or investigate malicious agent tool calls because no audit trail exists
Implement explicit audit logging for every MCP tool call: log the server identity, tool name, arguments \(with sensitive values redacted\), timestamp, and the LLM's stated reasoning for the call. Build real-time monitoring that alerts on anomalous patterns — unexpected tools, unusual call frequency, off-hours invocations, or calls to tools the agent rarely uses. Do not rely on the MCP protocol itself for observability; it provides none.
Journey Context:
The MCP protocol is designed for capability exchange and tool invocation, not for observability. There is no mandatory logging, no built-in audit trail, and no anomaly detection mechanism. If an agent is compromised via tool poisoning, prompt injection, or any other vector, its malicious tool calls execute silently with zero record. Developers assume that because tool calls happen through a structured protocol, they're automatically observable — they are not. The protocol transports the call and the response; what happens in between is a black box unless you instrument it yourself. Without explicit logging, you have no forensic evidence after an incident and no real-time detection during one. This gap is listed in the OWASP MCP Top 10 and is the most commonly overlooked operational control.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-15T01:34:09.438070+00:00— report_created — created