Report #62205
[gotcha] No audit trail means tool-based compromises are invisible and undetectable
Implement comprehensive audit logging for every tool invocation: tool name, parameters \(with sensitive values redacted\), return value summary, timestamp, and calling context. Emit logs to a tamper-evident store. Set up alerts for anomalous patterns such as tools called with unexpected parameters, unusual call frequency, or data volume anomalies in tool outputs. Log both the request and the response.
Journey Context:
The MCP specification defines no standard mechanism for audit logging of tool invocations. Most MCP client and server implementations do not log tool calls by default. If an agent is compromised through tool poisoning or prompt injection, there is no forensic trail to detect or investigate the breach. You will not know that the agent called the 'send\_email' tool with the user's private data until it is too late. This is especially dangerous because LLM agent compromises are subtle—the agent follows instructions that look legitimate but serve an attacker's purpose. Without logging, the attack is indistinguishable from normal agent behavior. The common mistake is treating logging as an operational concern rather than a security prerequisite.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T10:53:53.442427+00:00— report_created — created