Mistake Clinic
Module-Specific Mistake Radar
| Mistake | Signal | Repair |
|---|---|---|
| Benchmark equals safety | No system threats | Map assets and harms |
| LLM judge accepted blindly | Human disagreement unknown | Validate the evaluator |
| Prompt as security boundary | Tools overprivileged | Enforce authorization outside model |
| Red team without taxonomy | Anecdotes cannot regress | Version cases and mechanisms |
| Human fallback by slogan | No capacity or SLA | Test escalation operationally |
Practice Mistake Checks
Verify frozen sets, evaluator validity, threat model, privilege, privacy, affected groups, regression gates, stop control, and risk ownership.
Required evidence
Submit versioned tests, system manifest, raw and summarized results, representative failures, a rejected alternative, named control owners, and residual-risk decision. Safety claims without adversarial evidence do not pass.
Oral defense
Demonstrate one failure, identify its first violated boundary, explain why layered controls limit impact, and execute the stop or escalation path.
Source backbone
Use NIST AI RMF, OWASP LLM Top 10, NIST SSDF, and Building Secure and Reliable Systems.