Submitted to: National Institute of Standards and Technology (NIST), AI Risk Management Framework programme
Submission date: March 2026
Status: Submitted — frozen document; do not edit.
License: CC-BY-SA-4.0 (per AEGIS public-content convention)
What it is
A formal position statement responding to NIST’s AI Risk Management Framework, arguing that the AI RMF’s risk-management goals require runtime governance at the architectural layer in addition to the model-layer practices the framework currently emphasizes. The submission introduces the AEGIS architecture as a reference for how that runtime governance can be implemented, and grounds the argument in the empirical evidence the AEGIS edge laboratory has produced.
Position summary
- Risk management at the model layer is necessary but insufficient for autonomous agents that take action. RLHF, Constitutional AI training, and instruction tuning shape what models say; they do not constrain what agents do once acting.
- The action boundary is the correct enforcement point for the RMF’s
MEASUREandMANAGEfunctions when AI moves from advisory to autonomous deployment. - Architectural-layer governance is complementary, not competitive with model-layer alignment. Defense-in-depth framing should appear throughout NIST-aligned guidance.
- Reference artifacts are available for review: AEGIS Core (Apache 2.0, public), the ATX-1 threat taxonomy, and the empirical edge-laboratory evidence.
Status
The submission is on file with NIST and is one of two NIST-track engagements (the other is the NCCoE response on AI agent identity and authorization). Both are formally submitted, peer-validated, and treated as frozen documents in the AEGIS Initiative repositories — substantive edits would require a version bump and re-submission rather than amendment in place.
Canonical text
The authoritative submission lives at aegis-governance/docs/position-papers/nist/ in both Markdown and PDF form. The PDF is the version-of-record for citation; the Markdown is reference-only.
Relationship to other AEGIS work
- Argues for the architecture documented in the Cross-Cutting Runtime Enforcement paper.
- The empirical evidence cited in the submission is drawn from the AEGIS edge laboratory and is now Round 1’s public narrative.
- The threat-model component is grounded in ATX-1.
- The forthcoming Round 2 replication, when complete, retroactively strengthens this submission.
NIST policy context
Shapira et al.’s Agents of Chaos (2026) explicitly names NIST’s February 2026 AI Agent Standards Initiative (§16.5, p. 43) as the policy context for empirical agent-failure work. The AEGIS submission predates that initiative; the Round 2 replication, when complete, will be the artifact most relevant to the NIST initiative’s empirical-evidence needs. A separate follow-up note to the NIST AI RMF programme is planned for that point.