C5 for Prompt, Trace, and Model Logging

GenAI logging is a balancing act. Too little logging and you cannot investigate incidents or prove control operation. Too much raw logging and you create a sensitive data store full of prompts, documents, tool outputs, and personal data.

Log by Event Class

• Identity events: user, tenant, session, MFA status, agent identity
• Authorization events: requested action, resource ID, decision, reason, policy version
• Retrieval events: query metadata, candidate IDs, allowed IDs, denied IDs, classification
• Model events: provider, model, prompt template version, latency, error, safety result
• Tool events: API, method, scope, outcome, approval status, error path

Handle Payloads Carefully

Prompt text, retrieved chunks, and tool payloads should have stricter retention and access controls than metadata. Store full payloads only when the investigation, quality, or regulatory need is clear. Mask secrets and avoid storing tokens entirely.

Access Control for Logs

Logs from AI systems can contain the same sensitive data as production systems. Restrict access, require approval for sensitive trace inspection, log administrator access, and review permissions regularly.

Bottom Line

C5-aligned GenAI logging should reconstruct control decisions without turning observability into uncontrolled data replication. Metadata by default, sensitive payloads by exception, and strict access everywhere.