AI Engineer World's Fair 2025
Government Agents: AI Agents Meet Tough Regulations — Mark Myshatyn, Los Alamos National Laboratory
Loading slides…
Read the talk
Government Agents: AI Agents Meet Tough Regulations
Scientific agents can connect research papers to supercomputer simulations, but deploying them for federal missions also requires explainability, isolation, governance, and timely delivery.
From a talk by Mark Myshatyn
A chessboard without bishops
What is a nuclear science laboratory doing at an AI conference? Los Alamos’s answer begins with a chessboard. Mark Myshatyn, the laboratory’s enterprise AI architect, describes almost seventy years of applied AI and machine learning. In the photograph he presents, a scientist plays Los Alamos chess beside MANIAC I in 1956. The board has no bishops: Myshatyn connects the simplified game to a time when computer memory could not accommodate the full problem.
That photograph also belongs to the period after the Manhattan Project when the laboratory was advancing Monte Carlo methods still used today. Scientific computing is therefore familiar territory at Los Alamos. What agents add is a new opportunity to connect that accumulated mathematical machinery to systems that can gather information, propose work, and use scientific tools.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From papers to a simulated fusion capsule
The first demonstration turns that opportunity into a scientific workflow. Under pressure to deliver more mission work faster and at lower cost, the laboratory looks beyond what a model already knows to what the system can let it know. The task is to propose an inertial confinement fusion capsule for Lawrence Livermore National Laboratory. The workflow begins with literature, then reaches into the laboratory’s computing infrastructure.
The demonstrated progression is:
- Read an initial paper.
- Find and read papers related to it.
- Develop a capsule-design hypothesis.
- Execute generated code on high-performance computing assets to run thermodynamic and hydrodynamic tests.
The crucial transition is from proposing a design to testing it with scientific software. The LLM is only part of the system; the agent can use tools built from decades of mathematics, physics, and nuclear-stockpile stewardship work. Its output is not limited to a chat response containing code.
This creates a route to faster science while the risks that motivate the work are also accelerating. The demonstration ends with a simulation of a slice through an ICF capsule and a design the agent believed would optimize yield. That is a proposed optimization evaluated through simulation, not an experimentally established yield improvement.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Scaling beyond one scientific workflow
A capsule-design demonstration answers only a small part of the laboratory’s adoption problem. Myshatyn describes the broader setting as twenty thousand researchers, forty square miles of laboratories and test sites, and thirteen nuclear facilities. The National Security AI Office must advance AI science, including developing models internally, while recognizing that the laboratory cannot supply every capability itself. Commercial and academic partnerships are part of that operating model.
The same institution also needs payroll, procurement, and cybersecurity. Bringing generative AI into those workflows sits alongside the scientific mission, rather than outside it. Myshatyn points to academic relationships, including a partnership with the University of California family of schools, and acknowledged chemical and biological safety work with OpenAI. The laboratory contributes specialized data and experience conducting dangerous work safely—capabilities that remain valuable even to frontier labs able to recruit exceptional technical talent.
Hardware partnerships extend that collaboration into the computing environment. Myshatyn describes Venado, built with NVIDIA and HPE, as a system with over 2,500 Grace Hopper nodes. He reports bringing OpenAI models onto the system and into classified networks to work on problems specific to the laboratory’s data and mission. Access to early models or model weights depends on trust between partners; once those models become agents, that trust must also encompass responsibility for what the resulting tools and services do.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Faster adoption raises the stakes for providers
The policy setting is the administration transition in 2025 and the April 3 OMB memoranda M-25-21 and M-25-22, following the replacement of earlier executive orders. Myshatyn reads their direction as a push for faster government AI adoption, with an accompanying obligation to address real-world consequences. At Los Alamos, a data exposure can become a geopolitical or physical threat, not simply a commercial incident.
Mission adoption must go deeper than office productivity. Faster presentation creation or email summarization does not exhaust the opportunity. As AI moves further into mission workflows, customers entrust providers with more consequential data and outcomes. The capsule demonstration is an example of open, unrestricted science; controlled and classified workloads create a different relationship between customer trust and supplier responsibility.
Myshatyn emphasizes the enduring sensitivity of nuclear information when discussing Restricted Data and Formerly Restricted Data: the underlying physics does not expire. His description of information as born classified and remaining classified conveys the handling problem facing suppliers, rather than a blanket legal rule for every record in those categories. A provider must be able to support the customer’s actual information restrictions before its agent can become part of the mission.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The security requirements already exist
A supplier arriving with a SOC 2 report has not necessarily answered the laboratory’s security questions. Myshatyn’s reference point is NIST SP 800-53 Rev. 4, whose catalog he describes as containing more than a thousand controls and enhancements. The catalog’s size is not a claim that every deployment must implement every control; it illustrates the breadth of the government assurance system that a commercial attestation alone does not replace.
FedRAMP attempts to make cloud assurance more reusable through assessed control sets and continuous monitoring. Myshatyn uses two hundred, three hundred, and four hundred controls as illustrative assessment scopes, not a definitive mapping of baselines. He also describes a changing FedRAMP program placing greater emphasis on a provider’s continuing security posture. Trust has to persist after the initial assessment, especially when the service is producing agent outcomes as well as holding data.
Defense workloads add requirements tied to the information being handled. The DoD Cloud Computing Security Requirements Guide uses impact levels to shape service access for data such as personal information, mission information, operational information, and financial information. Myshatyn then adds CNSSI 1253 to the picture as another security-control layer. These frameworks are not interchangeable badges: the applicable requirements depend on the workload and its environment.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Developing AI governance with the customer
AI-specific governance is less settled than the existing cybersecurity requirements in the talk’s timeframe. Following the April memoranda, agencies face 180-day deadlines for strategies and compliance plans. These are implementation deadlines, rather than a general public rulemaking period. The practical questions include how to govern pilots, how to plan deployment, and what counts as high or low risk in a particular mission.
NIST’s AI Risk Management Framework 1.0, published in 2023, offers voluntary guidance; it is distinct from binding agency or procurement requirements. With implementation practices still being developed, providers have an opportunity to work with government customers on governance that preserves mission speed without obstructive process. That collaboration does not eliminate the records and artifacts that must exist. It can make those requirements easier to satisfy.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The science behind the paperwork
The reason to undertake that work is access to scientific problems and capabilities that are unusually difficult—or unavailable—in commercial settings. Myshatyn describes petabytes of laboratory data that have never been on the Internet and will never be placed there. The relevant expertise spans chemistry, biology, materials physics, composites, cybersecurity, and high-performance computing design.
Those capabilities do not make the laboratory self-sufficient. Protecting national competitive advantage requires engagement beyond the scientists at Los Alamos. The invitation to industry is to combine complementary expertise, data, and tools, with architecture that makes the partnership possible.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Four requirements for deployable agents
The closing architecture guidance starts with explainability. When an agent reaches a consequential decision, the customer needs to understand how it got there—especially after something goes wrong. For the laboratory, accountability runs to US citizens, not shareholders. Trusting an agent therefore requires the ability to investigate its decisions with the seriousness applied to the actions of human staff.
The second requirement is isolation. Los Alamos already relies heavily on self-hosted and open-source tools, services, and models because some necessary capabilities cannot be obtained from a hyperscaler offering. Myshatyn mentions looking forward to Microsoft’s self-hosted AI Foundry demonstration; he is not reporting an evaluation of it. The underlying requirement is an architecture that can operate where the customer’s data is allowed to reside.
Even a SaaS builder should inspect a cloud provider’s services-in-scope list. Myshatyn proposes the limited service set of a DoD Impact Level 5 environment as a useful design constraint: supporting that smaller set reduces dependencies and makes deployment easier for the customer. Treat this as a portability heuristic, not an assertion that IL5 support automatically authorizes deployment everywhere.
The third requirement is governance built into delivery. Myshatyn describes a software bill of materials as a required artifact in the laboratory’s procurement conversations. A surprised offer to dump a build script illustrates supplier unpreparedness, not an adequate SBOM process. Providers should be ready to explain how they manage open-source dependencies and how they plan to patch them. AI capabilities may advance faster than traditional cybersecurity processes, but helping a customer complete the necessary SaaS or PaaS documentation can make the partnership move faster.
The fourth requirement is speed in the constrained environment itself. Myshatyn recounts asking providers why federal offerings lag commercial ones, sometimes by a year, and raises one-, three-, and five-year parity delays in those conversations. These are examples of supplier discussions, not an industry benchmark. The problem is not simply a preference for newer features: export-compliance constraints and permitted deployment locations can make an otherwise available service impossible to buy. Designing release delivery for those difficult environments improves the chance that a tool can actually be fielded for the mission.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From national security to rocks on Mars
Los Alamos was founded on the premise that applying mathematics and science can change the world almost overnight. That history informs Myshatyn’s description of AI as both the greatest opportunity and the greatest threat to national security. The opportunity, he says, is what motivates participation: the laboratory needs to help develop the technology’s future despite its downside risks.
His final example reaches beyond the nuclear enterprise. Nuclear nonproliferation work helped the laboratory develop expertise in specialty sensors. Myshatyn connects that expertise to ChemCam, the Mars rover instrument Los Alamos helped build, which uses a laser to examine rocks. Capabilities developed for a sensitive mission can also advance fundamental science. The invitation to collaborate includes that broader possibility: tools built to address difficult security problems can expand what people are able to discover elsewhere.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
NIST's voluntary framework for managing risks when designing, developing, deploying, and using AI systems.
Further reading
The March 2025 proposal for automated cloud security assurance and continuous validation.
Research on combining reasoning agents, physics codes, and emulators to optimize fusion capsule geometry in simulation.
Updates since the talk
An August 2025 NNSA announcement describing Venado's hardware and operation of OpenAI reasoning models on a classified network.
Full transcript
Loading transcript…