Los Alamos is using agents to connect scientific literature, generated hypotheses, and high-performance computing simulations for problems such as inertial confinement fusion capsule design.
2
Government AI systems need controls for explanation, isolation, governance, and deployment speed because their data and outputs can affect national security and human lives.
3
Vendors that want to work with federal labs must plan for classified environments, continuous security monitoring, software bills of materials, export rules, and long procurement processes.
Summary
Mark Myshatyn explains why Los Alamos National Laboratory is attending an AI engineering conference. The lab has used applied statistics and machine learning for decades, but agents now let it connect models with its scientific software, high-performance computers, and specialized data. In one demonstration, an agent read papers, proposed an inertial confinement fusion capsule design, and ran thermodynamic and hydrodynamic tests on laboratory computing systems. The harder problem is deploying AI across a large national-security organization. Myshatyn describes layers of federal and Department of Defense security requirements, while AI-specific governance is still being written. He asks vendors to build systems that can explain decisions, operate in isolation, support governance artifacts, and run in restricted environments. Los Alamos also wants partnerships with commercial companies and universities because its offline data and scientific expertise create problems that commercial labs cannot easily reproduce. His position is direct: AI creates serious risks for national security, but avoiding the work would also leave the future to others.
Los Alamos has a long history of applied machine learning
Myshatyn places current AI work in Los Alamos's older scientific tradition. A 1956 photograph shows a scientist playing Los Alamos chess in front of the MANIAC I supercomputer, without bishops because the computer could not hold a full chessboard in memory. The lab was already developing Monte Carlo methods after the Manhattan Project, and it still uses them. He argues that generative AI and agents are new opportunities for a laboratory that has worked with applied statistics, simulation, and scientific computing for decades.
An agent can connect research papers to real scientific simulations
Los Alamos gave an agent a design problem for an inertial confinement fusion capsule and asked it to read the initial paper, find tangential papers, and propose a design. The agent formed a hypothesis and generated a design intended to improve yield. The important part was the connection to the lab's high-performance computing systems. The generated code ran thermodynamic and hydrodynamic tests, so the system combined a language model with decades of mathematical and scientific software rather than returning code from a chat window.
Myshatyn asks what the fusion demonstration means for Los Alamos's roughly 20,000 researchers, its large collection of laboratories and test sites, and its 13 nuclear facilities. The national security AI office has several jobs: advance AI science, write its own models and systems, and bring generative AI into ordinary workflows such as payroll, procurement, and cybersecurity. Los Alamos does not expect to do this alone. It needs commercial and academic partners, and it is already working with frontier labs on chemical and biological safety.
Government AI carries consequences that ordinary business software does not
Myshatyn describes recent OMB memoranda as pushing agencies to adopt AI while recognizing that government systems have real-world effects. A lab handling national-security data cannot treat AI as a faster way to create presentations or summarize email. If sensitive data escapes, or if an agent produces a bad outcome, the result can include geopolitical or kinetic consequences and people can die. The lab therefore needs trust in both the data handling and the outcomes produced by its agents.
Federal security requirements are layered and still changing
Vendors working with Los Alamos face requirements beyond a SOC 2 report. Myshatyn mentions NIST 800-53, FedRAMP, continuous monitoring, and the Department of Defense's cloud security requirements and impact levels. Additional requirements cover personally identifiable information, mission data, operational data, and financial data. AI governance remains under development while agencies write strategies, pilot rules, and definitions for high- and low-risk use cases. He points to NIST's 2023 AI Risk Management Framework as existing guidance.
Restricted environments create a shared design opportunity
Myshatyn says the government and vendors are working from a relatively open technology starting point for AI governance. Los Alamos has petabytes of data that have never been connected to the internet and will not be connected to it. The lab also has expertise in chemistry, biology, materials, physics, cybersecurity, and high-performance computing. Partnerships can give companies access to problems and knowledge that are difficult to reproduce commercially, while helping agencies write rules that support useful systems.
Federal agents need explanation, isolation, governance, and speed
Myshatyn closes with four design requirements for vendors bringing agentic tools to the federal government. Systems need to explain how they reached a decision, especially when an outcome causes public scrutiny. They need isolation because the lab handles sensitive and classified work. Vendors should support governance artifacts such as software bills of materials, dependency records, and patching plans. They also need to keep federal offerings current and design for the locations and export restrictions where government users must operate.
National-security work is also fundamental science
Myshatyn acknowledges that AI is both a major opportunity and a major threat to national security. He gives Los Alamos's ChemCam sensor as an example of work that began in the nuclear enterprise and later enabled a laser-equipped vehicle to analyze rocks on Mars. His point is that engaging with a national laboratory can lead to scientific applications beyond weapons work. Los Alamos wants outside help developing AI while accepting the risks that come with the mission.
"We have also had some awkward conversations with some of our service providers saying, why is your federal stuff a year out of date?"14:18
Who should watch
You build AI agents or SaaS products and want to sell to federal agencies or national laboratories.
Your system will handle classified, restricted, mission, or other sensitive data and you need to understand the deployment constraints.
You work on AI governance and want a concrete account of how security, explainability, procurement, and export rules affect scientific systems.
