Adversarial Red-Team Suite

Adversarial Red-Team Suite A workflow diagram generated by Archify. 01 / Threat taxonomy 02 / Layer 03 Evals & Red Teaming as Evidence 03 / Evidence EX / Accepted risk Build and version Run and triage Close and record Threat taxonomy · named techniques · Threat taxonomy › Build and version Threat taxonomy named techniques Red-team suite · versioned with model · Layer 03 Evals & Red Teaming as Evidence › Build and version Red-team suite versioned with model Run the suite · findings per technique · Layer 03 Evals & Red Teaming as Evidence › Run and triage Run the suite findings per technique Triage · fix / accept · Layer 03 Evals & Red Teaming as Evidence › Run and triage Triage fix / accept Fix shipped · regression case · Layer 03 Evals & Red Teaming as Evidence › Close and record Fix shipped regression case Evidence record · against registry · Evidence › Close and record Evidence record against registry Risk accepted · owner, rationale · Accepted risk › Close and record Risk accepted owner, rationale record record triage run derive accept fix Legend Agent logic Policy Context / trace External system

A standing adversary

  • • Cases are drawn from a threat taxonomy, so each traces to a named technique
  • • The suite is versioned alongside the model and run against the registered version
  • • The Eval Gate enforces a threshold; the suite finds the input it never anticipated

Every finding is closed

  • • Fix, or accept with a recorded owner and rationale
  • • Confirmed findings become regression cases, so a closed attack stays closed
  • • The outcome is filed as evidence against the registry entry

First step

  • • Treat red-team findings like test failures: an owner and a deadline
  • • Start with the techniques your system is exposed to today