Guarding tool calls

[wanyixu199]

Hi everyone, I’m collecting real-world security issues involving tool-using agents (LangGraph/LangChain). If you’ve seen an agent make an unsafe tool call or leak data in a surprising way, I’d love to hear the pattern (even high-level/redacted).

In return, I can share a tiny harness + guardrail approach that blocks risky tool calls and captures an evidence trail for debugging.