AARM Benchmarking Framework
Open Until: 10/30/2026
This document defines the AARM Benchmarking Framework: the structure of evaluation, the metrics, the scenario set covering the eight primary threats in the AARM specification, and the rubric by which results are scored.
Key Takeaways
- Real-World Threat Testing: Evaluate defenses against complex, multi-step attacks beyond basic conformance.
- Comprehensive Metrics: Balance security efficacy with operational friction by tracking detection and false-positive rates.
- Full Threat Coverage: Assess robustness across eight core categories, including prompt injection, data exfiltration, and intent drift.
- Four-Tier Rubric: Measure progress from baseline compliance to production-grade enterprise readiness.
