Report #91527
[gotcha] Agent silently executes a malicious tool action without logging the invocation payload or result making forensics impossible
Implement mandatory, immutable audit logging for every tool invocation \(input, output, timestamp, and the LLM's reasoning for calling it\) outside the agent's control.
Journey Context:
Developers log LLM prompts but forget to log the actual side effects. If an agent is socially engineered into deleting a database, and the \`delete\_db\` tool doesn't log who called it and why, you have a breach with no root cause. The agent's 'thought' process is ephemeral; tool execution is permanent. Telemetry must be decoupled from the agent so a compromised agent cannot wipe its own logs.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T12:13:12.301588+00:00— report_created — created