Government Agents: AI Agents vs Tough Regulations — Mark Myshatyn, Los Alamos National Laboratory

Los Alamos National Laboratory (LANL) has a history of applied AI and machine learning stretching back nearly 70 years, beginning with running chess on the Maniac 1 supercomputer in 1956 with such limited memory that the game had to be simplified. This legacy of pushing computational boundaries, including the development of Monte Carlo methods, provides the foundation for the lab’s current work in the age of generative AI and agentic systems.

Accelerating Science with Agentic AI

The lab views generative AI not just as a model but as an agentic capability to accelerate science. A key demonstration involved tasking an AI agent to design an inertial confinement fusion (ICF) capsule. The process involved:

1. Ingesting Knowledge: The agent began by reading a foundational scientific paper on the topic, then independently found and read tangential papers to build a comprehensive understanding.
2. Hypothesis and Design: It formulated a hypothesis and created a novel design for a fusion capsule.
3. Execution on HPC: Critically, this was not a theoretical exercise. The agent generated code that was then executed on LANL’s high-performance computing (HPC) assets, running complex thermodynamic and hydrodynamic simulations to test the design’s efficacy and optimize for yield.

This approach integrates modern LLMs with LANL’s decades of specialized scientific software and simulation tools, creating a powerful system where an AI can actively participate in the scientific discovery process.

LANL’s National Security AI Strategy

The laboratory’s AI office operates on three core principles:

1. Advance AI Science: Proactively push the frontiers of AI by writing their own models and tools rather than only consuming commercial or open-source offerings.
2. Foster Partnerships: Recognizing the complexity of the field, LANL actively partners with commercial industry (e.g., OpenAI for chem-biosafety research, Nvidia and HPE for hardware) and academia (e.g., the University of California system) to co-develop the future of AI.
3. Operationalize AI: Integrate AI and generative AI tools into the lab’s vast internal workflows, including procurement, payroll, and cybersecurity, to improve efficiency across its 40-square-mile campus.

A cornerstone of this strategy is the Venado supercomputer, an HPE system with over 2,500 Nvidia Grace Hopper superchips. LANL has deployed frontier models like those from OpenAI onto this system within its classified networks, enabling research on sensitive, mission-critical data that cannot be exposed to the public internet.

The Imperative of Trust and Governance in Federal AI

Deploying AI in the national security context carries immense responsibility, where failures can have geopolitical or kinetic consequences. The U.S. government is formalizing its approach to AI governance, notably through the Office of Management and Budget (OMB) memorandum M-24-10, which directs agencies to accelerate AI adoption while managing its risks.

For companies wishing to partner with the federal government, this translates into a complex but non-negotiable set of security and compliance requirements. This shared responsibility model requires vendors to navigate a landscape of rigorous standards:

• NIST 800-53: A framework with over a thousand security controls.
• FedRAMP: The baseline authorization process for cloud services in the government, which is evolving to require more robust continuous monitoring.
• DoD Security Requirements Guide (SRG): Imposes additional “Impact Levels” (e.g., IL5, IL6) for handling more sensitive data.
• CNSSI 1253: An additional layer of controls for national security systems.

While this governance landscape is intensive, the rules for AI-specific use cases are still being written. This presents a unique “clean sheet of paper” opportunity for commercial partners to collaborate with agencies like LANL to develop better, less obstructive, and more effective governance frameworks for the future.

An Architectural Blueprint for Government Collaboration

To successfully partner with federal agencies in high-stakes environments, vendors should build their tools and services on four key architectural pillars:

1. Build for Explainability: In a government context, accountability is to citizens, not shareholders. Agents must be trustworthy, and their decision-making processes must be auditable, especially when an outcome is negative.
2. Build for Isolation: LANL operates with petabytes of data that will never see the internet. Tools must be designed to function in air-gapped or highly restricted environments. A key piece of advice is to build for a DoD Impact Level 5 (IL5) environment; if a tool can operate with the limited services available there, it has achieved a “least common denominator” portability and can be deployed almost anywhere.
3. Build for Governance: Be prepared to provide a Software Bill of Materials (SBOM) and have transparent processes for managing open-source dependencies and patching. This is a standard cybersecurity requirement that is essential for procurement.
4. Keep Up the Speed: Federal offerings from cloud and software vendors often lag years behind their commercial counterparts. This gap is not sustainable and is often rooted in export compliance laws. Designing for speed and feature parity even in the most restricted environments is crucial for a successful partnership.

Ultimately, LANL was founded on the idea that math and science can change the world. The lab views AI as the greatest opportunity—and threat—to national security, but it is the opportunity that drives its mission to help invent the future.

• Tags:
• Agentic AI
• AI Governance
• NIST
• Feature
• Los Alamos
• HPC
• National Security
• FedRAMP
• Scientific AI