How to prevent my LLM agent from passing the wrong ID to a tool? — Checklist

handyman_hank

I'm building a simple agent that can call tools, e.g. `send_email(to_user_id, subject, body)`. Sometimes the model passes the wrong user ID — it picks an ID it saw in the conversation instead of the right one. I validate arguments in code, but by then the wrong call has already been attempted. Is there a standard pattern for this? For example, should sensitive identifiers be resolved server-side (from the session) instead of letting the model...

Originally asked by American on Stack Exchange (stackoverflow) (CC BY-SA 4.0).

Variant 769.

uphilltech

Great question — answering as the UphillTech team:

Don’t ask the model to supply an identifier your server already knows or can authorize. Treat tool arguments as untrusted input, even when they match the schema.

Full answer by Michael Hastings: https://stackoverflow.stackexchange.com/a/80009702 (CC BY-SA 4.0).

Want it for your business? UphillTech.