Orithos Internal Red-Team — Evaluation Topology

Orithos Internal Red-Team — Evaluation Topology An architecture diagram generated by Archify. Orithos Console · scan configuration · Orithos platform (orithos.com) · operator Orithos Console scan configuration operator Scan API · POST /v1/scans · Orithos platform (orithos.com) · apps/api Scan API POST /v1/scans apps/api ARQ Queue · run_scan_job · Orithos platform (orithos.com) · dispatch ARQ Queue run_scan_job dispatch Scan Worker · async execution · Orithos platform (orithos.com) · workers/scan-worker Scan Worker async execution workers/scan-worker Probe Engine · 1,092 probe keys · Orithos platform (orithos.com) · httpx sessions Probe Engine 1,092 probe keys httpx sessions AEGIS-J Evaluator · V-FULL…V-NONE verdicts · Orithos platform (orithos.com) · cascade judge AEGIS-J Evaluator V-FULL…V-NONE verdicts cascade judge PostgreSQL · findings + traces (RLS) · Orithos platform (orithos.com) · evidence store PostgreSQL findings + traces (RLS) evidence store OpenRouter Gateway · openai-compatible · evaluation targets · model access OpenRouter Gateway openai-compatible model access 6 Demo Personas · Meridian · NexusHealth · Fortis +3 · evaluation targets · internal targets 6 Demo Personas Meridian · NexusHealth · Fortis +3 internal targets create scan enqueue dequeue probe batches response + trace verdict + confidence attack payloads completion calls Orithos platform (orithos.com) evaluation targets Legend Frontend Backend Database Cloud Security Message bus External

How the evaluation ran

  • • 126 scans dispatched from the Orithos console against internal targets
  • • Probe engine drove 2,000 evaluated outcomes through the model gateway
  • • Every outcome persisted with trace and verdict — nothing inferred after the fact

Boundary honesty

  • • Targets were our own demo personas, not customer agents
  • • Gateway traffic is isolated per probe session — one agent, one client
  • • Findings carry evaluator verdicts, not heuristic flags

Verified by this dataset

  • • 1,092 unique probe keys exercised across 126 scans
  • • 80 scans completed; 43 failed and 3 aborted — disclosed
  • • Verdicts: 85 V-FULL, 66 V-PARTIAL, 26 V-INDIRECT, 26 V-LEAK