How to detect unauthorized AI agent actions in production

Span-based instrumentation and activity schemas surface prompt injection, unauthorized access, and multi-turn drift before incidents compound.