COMPASS
Security maturity / scoring (threat prioritization workflow)OWASP Threat Defense COMPASS · OWASP GenAI Security Project
COMPASS supplies a fast OODA-loop threat-prioritization workflow that ranks known AI threats by Impact x Likelihood. AITBM can complement that workflow with multi-dimensional, confidence-graded system assessment; a COMPASS threat-row priority is not numerically interchangeable with an ERS.
| COMPASS Dimension / Threat Class | Primary AITBM Sub-Metrics | Evidence Use / Notes |
|---|---|---|
| Impact (1-5) dimension | IVP sub-metric severity + ORP Cp | Input construct; COMPASS blends failure severity with blast radius that AITBM separates |
| Likelihood (1-5) dimension | IVP sub-metric exposure + ORP As | Input construct; maps to exploitability and deployment exposure |
| Prompt injection (LLM01) | Ro-1Cn-1 | Adversarial ASR and unauthorized-action evidence; no generic ERS |
| Sensitive disclosure (LLM02) | Pr-1Pr-4 | Membership-inference and leakage evidence; no generic ERS |
| Excessive agency (LLM03:2026) | Cn-1Cn-2Cn-5Cn-6 | Unauthorized-action, identity, and reversibility evidence; no generic ERS |
| Misinformation / hallucination (LLM07:2026) | Ro-3Tr-2 | Hallucination-rate and calibration evidence; no generic ERS |
| Bias / discriminatory output | Fa-1Fa-3Fa-4 | Demographic-parity and counterfactual-fairness evidence; no generic ERS |
| Agent impersonation / multi-agent trust | Cn-5 | ISSR and MTTQ evidence; no generic ERS |
| OODA cadence (continuous re-run) | ACI Tf (Temporal Freshness) | A qualifying re-run may refresh Tf for covered evidence; AITBM admissibility and event rules govern |
Key findings
- COMPASS scores individual threat rows on two assessor-estimated 1-5 scales (Impact and Likelihood). In a combined workflow, AITBM supplements that priority cell with a system-level 0-10 ERS and preserved per-axis profile; it does not replace COMPASS's threat-prioritization output.
- A COMPASS Impact x Likelihood cell entangles failure severity, deployment context, and confidence; AITBM separates these into IVP, ORP/CRM, and ACI so remediation can target the weakest axis (e.g. Cn-1, Cn-5) rather than an opaque '4x4'.
- There is no priority-to-ERS numeric crosswalk: COMPASS ranks one threat, ERS scores a whole system. The integration is evidence flow (score each row's sub-metric -> compose to ERS) and writing ERS-derived severity back into COMPASS.
- The COMPASS worked example is retained on its dated 21-sub-metric, pre-GDCP basis. A current assessment must derive Cn-6, Cp, ACI, and ERS under the current specification.