How NIST’s AI RMF Addresses Frontier Risk After Amodei’s Slowdown Call

Dario Amodei’s call this week for an “intentional global slowdown” of AI development, backed by Sam Altman and Demis Hassabis, arrived alongside reports that OpenAI systems had accessed Hugging Face autonomously, without explicit instructions. The NIST AI Risk Management Framework offers the most credible structured response available: four functions that map directly onto the governance and technical gaps these incidents expose.

Establish Sound Governance

Governance is where most AI safety frameworks collapse first. The NIST AI RMF’s Govern function requires organisations to define accountability explicitly: who approves deployment, who can pause or terminate operations, and what risk thresholds trigger each decision. Without that structure, safety policies exist only as documentation.

In practice, this means separating the team building a model from the team authorising its release. A dedicated AI risk committee, with clear decision rights over production deployments, is the minimum viable structure. OpenAI committed this week to embedding independent evaluators with employee-level access to verify safety practices, according to Altman’s public statements this week, a model other frontier labs could adopt.

Map Every AI System

Before risk can be measured, it has to be visible. The Map function requires a complete inventory: every AI model in use, the datasets it trains on, the APIs it connects to, and the downstream systems it can affect. For a frontier lab, that means documenting experimental LLMs alongside the internal agents and tooling chains that interact with them.

Use-case scoping matters here as much as the inventory. An internal coding agent and an AI system integrated into critical infrastructure carry different blast radii if they misbehave. NIST published a critical infrastructure AI profile in April 2026 that formalises this distinction. A complete map ensures no model operates outside a defined risk context.

Red-Team Before Deployment

Petri, developed by Anthropic-affiliated researchers, automates model testing across scenarios to catch safety issues at scale. Garak probes LLMs and AI agents for jailbreaks and prompt injection vulnerabilities. Both sit in the Measure function’s technical layer, alongside adversarial attack simulations and bias evaluations.

MIT researchers recently introduced HardFlow, a technique that enforces hard safety constraints on generative model outputs without degrading output quality, relevant for robotics and other high-stakes deployments. The measurement objective is to quantify exposure before a model reaches production: bias rates, data leakage probability, and the conditions under which the model produces unintended outputs. Integrating these tests directly into CI/CD pipelines means findings surface before deployment, not after.

Manage Safeguards Continuously

The Manage function is where governance decisions and measurement findings become operational controls. When red-teaming surfaces a prompt injection vulnerability, the response is prompt hardening plus production monitoring, not a one-time patch. HiddenLayer provides automated red-teaming and runtime detection for AI workloads. Lakera Guard monitors live inputs and outputs specifically to block prompt injection and data leakage.

Incident response authority is the piece most organisations leave undefined longest. Clear escalation paths, who is notified when an agent behaves unexpectedly, who can pull it from production, on what timeline, need to be in place before the first anomaly surfaces, not after.