AI Governance for Defense Contractors: NIST AI RMF Compliance Through Structured Delegation
As a systems engineer with 25 years in the defense sector, I’ve seen firsthand the increasing pressure to adopt AI—and the equal pressure to do so responsibly. My current work building ARKONA, an autonomous multi-agent AI ecosystem, is forcing a deep exploration of AI governance, particularly regarding NIST AI Risk Management Framework (RMF) compliance. The challenge isn’t just *using* AI, it’s demonstrating accountability and mitigating risk in highly regulated environments. I’m sharing insights gleaned from ARKONA’s architecture, specifically how we’ve implemented a structured delegation framework to address these concerns.
The Problem: Operationalizing the NIST AI RMF
The NIST AI RMF is a fantastic document, outlining crucial steps to identify, assess, and manage risks associated with AI systems. However, moving from the conceptual framework to a practical implementation for a complex, distributed system like ARKONA is non-trivial. Traditional risk management often relies on manual review and sign-off, which simply doesn’t scale with 26 autonomous agents operating on a battle rhythm, across 47 services running on encrypted internal network, spanning 23 ports. We needed automation, auditability, and a clear chain of responsibility.
Our initial approach involved mapping the four functions of the NIST RMF – Govern, Map, Measure, and Manage – to specific services within ARKONA. But that felt… static. The real power, we realized, came from *dynamically* applying risk mitigation through a delegation system. This is where our COMET framework came into play.
COMET: A 7-Step Human-AI Delegation Framework
COMET (Cyber Operational Management & Ethical Tasking) isn’t just a name; it represents our 7-step process for delegating tasks to AI agents while maintaining human oversight and accountability. It’s grounded in IEEE and NIST standards relating to autonomous systems, and it directly addresses several core tenets of the NIST AI RMF, particularly those around trustworthiness and responsible AI.
- Define Intent: A human operator (e.g., a Cyber Threat Intelligence analyst) clearly defines the task and associated risk tolerance.
- Agent Selection: The Inter-agent Communication Broker selects the most appropriate agent based on capability and current load. We leverage a dynamic capability model, updated via our DevOps pipeline, to ensure accurate selection.
- Risk Assessment: The NIST 800-30 grounded risk evaluation engine assesses the task against pre-defined risk profiles.
- Constraints & Guardrails: The human operator sets constraints – boundaries within which the agent can operate. These constraints are crucial for preventing unintended consequences.
- Execution & Monitoring: The agent executes the task. Real-time monitoring is handled by a dedicated monitoring agent, feeding data back to the operator.
- Provenance & Signing: All agent actions are cryptographically signed with SHA-256, establishing a clear provenance trail. This is fundamental for auditability.
- Review & Refinement: A human operator reviews the agent’s output, validates its accuracy, and provides feedback to refine the agent’s performance and improve future task delegation.
This isn't just a workflow; it's an architectural pattern. Each step is implemented as a microservice within ARKONA, communicating via our pub/sub messaging system. For instance, the risk assessment step calls our risk evaluation engine, passing the task description and agent capabilities as parameters. The engine returns a risk score and a list of recommended mitigation strategies, which are then presented to the operator.
Technical Implementation: Task Delegation via MCP
The core of COMET’s operation is the MCP (Mission Control Protocol) server, our inter-agent communication broker. It handles task delegation, status updates, and data exchange between agents. The MCP server utilizes a JSON-RPC based API, allowing for structured communication and easy integration with other services. Here's an example of a task delegation request:
{
"task_id": "CTI-20260407-001",
"agent_type": "HardwareRE",
"task_description": "Reverse engineer malware sample SHA256: a1b2c3d4e5f67890",
"constraints": {
"max_execution_time": 3600,
"allowed_domains": ["analysis.internal"]
},
"risk_profile": "High",
"operator_id": "JArango"
}
This request is routed through the MCP server to the hardware reverse-engineering service (integrated with Ghidra). The RE service then begins the analysis, adhering to the specified constraints. Crucially, the task ID and operator ID are logged throughout the process, providing a complete audit trail.
Newsroom Editorial Pipeline & Fact-Checking
Our 5-agent newsroom editorial pipeline exemplifies how we've extended COMET beyond technical analysis. This pipeline automates the process of gathering, summarizing, and publishing threat intelligence reports. However, it doesn't just publish; it includes a dedicated fact-checking agent. This agent, leveraging our MuXD hybrid LLM router (Ollama local + Claude cloud), validates claims against trusted sources. This is a direct application of the ‘Measure’ function of the NIST RMF – quantifying the accuracy and reliability of AI-generated content. The local Ollama models provide faster, cheaper analysis of common data points, while Claude handles more complex reasoning and contextual understanding.
We’ve designed the pipeline to trigger alerts if the fact-checking agent identifies discrepancies, automatically flagging the report for human review. This blended approach – automated analysis with human oversight – is essential for maintaining trust and ensuring the integrity of our intelligence products.
Provenance and Auditability: SHA-256 Signing
As I mentioned, all agent actions are signed with SHA-256. This isn't just about detecting tampering; it's about establishing a chain of custody. Every log entry, every data transformation, every analysis result is linked back to the responsible agent and operator. This is accomplished through a dedicated signing service that intercepts all inter-agent communication and applies the cryptographic signature. We store these signatures alongside the associated data, allowing for verifiable audit trails.
Currently, we’re seeing roughly 184 commits per week into ARKONA, demonstrating the rapid evolution of this ecosystem. Maintaining provenance is critical to understanding the impact of these changes and ensuring that our AI systems remain trustworthy.
Lessons Learned
Building ARKONA has reinforced a simple but profound lesson: AI governance isn’t about *preventing* AI adoption; it’s about *enabling* responsible innovation. The NIST AI RMF provides a valuable framework, but it needs to be operationalized through architectural patterns like structured delegation. Simply having a policy document isn’t enough. You need systems that enforce those policies automatically, provide clear audit trails, and empower humans to maintain meaningful control. By focusing on accountability and transparency, we can unlock the full potential of AI while mitigating the inherent risks.