Report #79853
[gotcha] MCP client logs tool invocation but redacts or omits the actual arguments passed by the LLM
Ensure the MCP client logs the complete, unredacted \`arguments\` object of \`tools/call\` requests and the \`content\` of the responses in a secure, append-only audit log. Do not rely on LLM reasoning traces \(like ReAct\) as the source of truth for what was executed.
Journey Context:
When an agent goes rogue, incident responders need to know exactly what data was exfiltrated or what commands were run. Clients often log 'Called tool: execute\_code' but drop the arguments because they can be large or contain PII. This makes forensic analysis impossible, as you cannot determine if the LLM injected malicious payloads into the arguments.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T16:38:30.384096+00:00— report_created — created