AI Agent Logs Hide Who Really Approved Actions
Who approved that AI action? The logs won't say — and that's a problem.
Imagine you're reviewing an AI assistant's actions after the fact. You see it made a tool call, but you have no idea if a human approved it or if an automated policy quietly waved it through. That's the problem a developer recently exposed in two popular open-source tools. When safety systems have workarounds — like allowlists, confidence thresholds, or special exception modes — those overrides usually aren't logged. So the trace logs look clean, but the truth is hidden.
Take retire.js, a tool that many developers use to scan for known security vulnerabilities. It has a flag called --insecure that disables a key security check, yet the final report looks exactly the same whether that flag was used or not. The developer reported this, and the maintainer agreed that it should be recorded — but fixing the logging didn't change the fact that the insecure option still exists and still works with no warning.
Langfuse, a platform for tracking AI agents, had a similar issue. Its built-in agent could mark tool calls as auto-approved based on a prefix, allowlist, or policy — but the code had no logging at all. A pull request added visibility, so each tool call now records whether approval came from a human, a conversation, or a policy. The fix was merged in 19 minutes, but the auto-approval rules themselves were left unchanged.
The developer's big worry: those bypass mechanisms are a permanent part of how these systems work, but they're rarely tested or reported as a whole. The fix shows you what actually happened, but it doesn't make the AI more cautious. It's like adding a speedometer to a car that still has no speed limit.