Your agent reads logs written by your users, your upstreams, and anyone who can make your service print a line. Treat every log line as untrusted input and clean it before it reaches the model.
Your Log File Can Hijack Your Agent: Sanitize Logs Before the Model Reads Them
