3

Exemplar Use Cases of Foundation Models

DEPARTMENT OF ENERGY’S ROLE IN FOUNDATION MODEL DEVELOPMENT

The strategic focus of a Department of Energy (DOE)-wide foundation model initiative remains a subject of debate, requiring the department to balance its broad application space, navigate the trade-offs between leveraging past industry advancements, address the unique national security imperatives of DOE, and ensure responsible stewardship of taxpayer resources, particularly in light of the opportunity costs associated with prioritizing artificial intelligence (AI) over more mature technologies.

There are an ever-increasing number of efforts across DOE national laboratories integrating AI and foundation models into their research programs. Naturally, one of the key targets is energy-related applications ranging from electric grids to nuclear fusion. A primary consideration is the perception that DOE cannot compete with the head start in technology maturation and large market share currently held by large companies. In the foundation model market, leaders such as Microsoft (via OpenAI), Google/Gemini, Amazon Web Services, Meta, and Anthropic each back efforts with investments ranging from $10 billion to more than $75 billion in funding and infrastructure (Fernandez et al. 2025), a scale that DOE would be hard pressed to match. This raises a fundamental question of whether DOE should focus on collaborations with industry or focus on a complementary space based on DOE’s unique mission in curating foundational science while improving national security. The committee believes that DOE has reason to develop foundation models internally, in addition to private-sector leadership, because the needs of the government (whether for national security or continued scientific preeminence) will not be met by private interests. The two

endeavors (private and public) do not compete—they complement each other and can leverage each other.

Conclusion 3-1: Commercial industry has driven rapid progress in developing large language model–based foundation models, yielding a robust ecosystem of tools and capabilities. As demonstrated by, for example, the collaboration between Los Alamos National Laboratory and OpenAI, DOE can leverage industry advances, findings, and collaborations as it develops foundation models for science and conducts coordinated DOE-wide assessments to identify appropriate opportunities.

DOE is the largest single federal sponsor of scientific research in the United States, providing approximately $16 billion in research and development (R&D) funding in fiscal year (FY) 2023, which represents roughly 8 percent of total federal R&D obligations (Blevins 2022). Through its Office of Science, DOE supports approximately 40 percent of all federal basic research in the physical sciences, and an estimated 44 percent of federal basic research in computer and information sciences, including foundational work in nonconvex optimization, probabilistic methods, and large-scale high-performance computing (NCSES 2023). Although DOE is unlikely to match the pace or scale of commercial product development, it retains clear strategic advantages in five areas: (1) a world-class scientific workforce in computational science; (2) access to large-scale, science-focused, and experimental computing hardware; (3) stewardship of unique experimental facilities and open and controlled or classified scientific data; (4) capability to tackle long-term, high-risk, high-reward scientific problems; and (5) access to unique scientific data that may not be easily reproduced and which can be expanded as synthetic data may be necessary for training future foundation models.

Despite a mismatch in funding allocations, DOE’s Exascale Computing Project (ECP)¹ guided the development, procurement, and construction of the Frontier and Aurora supercomputers at a total cost of approximately $1.7 billion. Frontier achieves a peak performance of 1.35 exaflops, and Aurora reaches approximately 1.01 exaflops. For a rough comparison, a machine like Aurora could train a model like GPT-4 on the order of ~200 days, suggesting that the best option for DOE is not to directly compete in the same general-purpose paradigm. With recent attention toward the disruption of DeepSeek, which some analyses suggest offered a ~10× increase in efficiency, existing ECP-funded resources become arguably more competitive, particularly when buoyed by the highly skilled workforce represented by the national laboratories. In fact, ECP-funded resources have the potential to train foundation models from scratch, deploy stochastic optimization algorithms at scale, or run multiagent simulations in real time. This is evidence that the field is advancing in a direction that could make DOE’s resources feasible for the training of foundation models for science.

___________________

¹ See https://www.exascaleproject.org.

When comparing DOE and industry capabilities, one of the most significant differences is the scale and richness of physical and simulation data generated by DOE’s network of user facilities, nuclear weapons–testing archives, ongoing experimental campaigns, and high-performance computing facilities. These data span both classified and unclassified domains and present unique opportunities for DOE-relevant advances in foundation models. New modes of inquiry that have become successful in the industrial setting (e.g., Google DeepMind’s alpha-evolve) have great potential for DOE applications. A central technical challenge is whether secure training methods, such as federated learning, can be designed to mathematically preclude the leakage of sensitive or controlled information. If so, this could enable the construction of scientific foundation models that operate across heterogeneous and compartmentalized data sources. If not, certain classes of model architectures may prove fundamentally incompatible with DOE’s mission constraints.

Conclusion 3-2: DOE retains clear strategic advantages in five areas: (1) a world-class scientific workforce in computational science; (2) access to large-scale, science-focused, and experimental computing hardware; (3) stewardship of unique experimental facilities and open and controlled or classified scientific data; (4) capability to tackle long-term, high-risk, high-reward scientific problems; and (5) access to unique scientific data that may not be easily reproduced and which can be expanded as synthetic data may be necessary for training future foundation models.

Many of DOE’s experimental platforms are already compatible with remote operation and automation. This includes user-facing beamlines, additive manufacturing facilities, and autonomous platforms for chemical synthesis and materials fabrication. At the same time, recent advances in retrieval-augmented generation (RAG) have introduced new strategies for connecting large language model (LLM) outputs with authoritative external sources. DOE could consider a coordinated program, either independently or in partnership with academia and industry, in which traditional physics-based simulations or experiments are launched in an agentic loop and used to refine LLM reasoning. This concept is particularly viable in diverse domains such as small-molecule chemistry and mature simulation codes (e.g., computational fluid dynamics, electromagnetism, molecular dynamics). An important example is the recent effort by researchers at Lawrence Livermore National Laboratory to combine AI with fusion target design by deploying AI agents on two of the world’s most powerful supercomputers to automate inertial confinement fusion simulations and thus accelerate experiments. Additional potential benefits of AI in the quest for fusion energy are provided in the next section.

However, in many scientific contexts, human expertise remains essential for initiating, interpreting, and validating results. Discovery via the use of experimental

or computational platforms relies crucially on the deep bench of technical expertise at the laboratories that can be rapidly tapped to analyze previously unseen scenarios in high-consequence national security settings with limited time to solution. In these settings, foundation models may act as an accelerant for analysis, but are currently not viewed as sufficiently reliable for trustworthy application.

HUMAN IN THE LOOP AND ARTIFICIAL INTELLIGENCE FOUNDATION MODEL AUTONOMY

Human oversight remains essential in deploying and utilizing foundation models, especially in high-risk or high-impact scientific and engineering contexts. Foundation models enhance the productivity of researchers by, for example, accelerating targeted literature reviews, optimizing code and algorithm design, and dramatically reducing the time required to prototype and validate solutions. By automating many of the routine or well-established steps in the research process, foundation models allow scientists and engineers to focus on higher-level reasoning and innovation. However, it is important to keep in mind that these capabilities come with a significant caveat: foundation models are capable of generating both highly sophisticated statements and nonsense. Although they can produce novel findings and accurate solutions, they can just as easily generate plausible-sounding but incorrect or misleading outputs. For this reason, and for the foreseeable future, a human-in-the-loop approach is desirable (even essential), ensuring that domain expertise and critical thinking guide the use and interpretation of model outputs. In this context, we mention that there are different levels or schemas of handoff, that is, the transfer of decision-making authority or control between a human and a foundation model (or agentic environment). The nature of the handoff depends on both the confidence in the model’s output as well as on the level of criticality (risk) associated with the decision one is trying to make. Importantly, this concept and associated schemas will evolve as foundation models become more mature and trustable (see Chapter 5 for details on quantifiable confidence).

A powerful example of this human–AI interaction is DeepMind’s AI co-scientist work (Gottweis and Natarajan 2025), a multiagent system built on Gemini, designed to assist scientists, engineers, and researchers in general in formulating hypotheses, conducting literature reviews, and building experimental frameworks. In this work, specialized agents operate asynchronously to generate, evaluate, and fine-tune scientific hypotheses. In several instances, this collaborative approach has made it possible for scientists to interact easily and very naturally with AI, providing inputs, prompts, or feedback to guide research, with final oversight and authority remaining with the investigator. For example, the AI Co-Scientist has demonstrated its potential impact in biomedical research, suggesting novel approaches to inhibit disease progression in conditions, such as liver fibrosis, that showed promising potential.

Another compelling example of human–AI interaction enabled by foundation models is the use of AI copilots in software development. Importantly, this approach is quickly becoming the norm in modern software engineering. For example, tools such as Cursor, which is built on top of LLMs and fine-tuned for code generation, offer real-time support in writing, debugging, and refactoring code (Anysphere n.d.). These systems serve as smart and efficient collaborators, helping developers implement complex algorithms and explore alternative code-design patterns. Some of these tools integrate seamlessly with developer workflows, allowing users to query codebases in natural language, generate multifile implementations, and suggest algorithmic solutions from minimal user prompts and/or examples. Although these tools do not replace developers, they act as accelerators, cutting down on repetitive and established coding tasks. As a result, engineers can focus on architectural decisions and problem solving. Notably, this also lowers software skill requirements. Again, the human remains in the loop; although the model may generate functional code, oversight is mandatory to validate correctness.

This paradigm illustrates how the synergy between human expertise and foundation model capabilities can lead to more efficient, reliable, and responsible scientific outcomes.

Conclusion 3-3: While AI systems can exceed human performance in many ways, they can also fail in ways a human likely never would. For this reason, the qualification of AI will be necessary for decision making and prediction in the presence of uncertainty.

Based on the discussion above, the integration of foundation models into scientific and engineering pipelines raises concerns about the future of employees working in these sectors. In fact, although these models boost productivity, they put at risk those roles that are focused on repetitive, manual, or routine tasks. Roles dedicated to basic literature reviews, boilerplate coding, standard documentation, and straightforward data analysis could be significantly affected. Importantly, in many cases, foundation models could be able to perform at scale and with more reliability than that of a human.

On the other hand, roles that require deep domain expertise and critical judgment (e.g., principal investigators, senior engineers, code architects, and regulatory or quality assurance engineers) are less likely to be removed. In fact, because of the need for human oversight when it comes to the interpretation of output of foundation models, these roles become even more valuable, as they are fundamental in verifying and building on top of what AI-based machines can achieve.

In short, foundation models potentially introduce a paradigm shift, where humans act as big-picture strategists and critical evaluators of AI-generated outputs, ensuring that they are technically correct and aligned with larger scientific and engineering goals.

Recommendation 3-1: The Department of Energy (DOE) should study and develop the fusion of artificial intelligence (AI) and human capabilities. At present, AI systems handle the repetitive, manual, or routine tasks, and are starting to show abilities to reason. As AI becomes more capable, deep analysis and strategy recommendations become feasible, but humans should maintain oversight and validation, particularly for qualification and other aspects of DOE’s mission.

Recommendation 3-2: The Department of Energy should evaluate the capabilities and risks of agentic artificial intelligence (AI) systems for its core applications. In particular, the committee advocates exploring agentic AI for developing autonomous laboratories for scientific discovery, decision making, and action planning for high-stakes applications.

SCIENTIFIC AND ENGINEERING APPLICATIONS

In the context of scientific and engineering applications, foundation models (FMs) trained on observations, scientific literature, databases, as well as experimental results and outputs of simulations can be used to support hypothesis generation. In engineering, the use of FMs is becoming predominant in design settings, specifically in tasks such as CAD (computer-aided design) generation. These applications demonstrate how FMs can serve as intelligent copilots for researchers and engineers, enhancing productivity and enabling new modes of discovery.

DOE’s mission encompasses many areas including materials science, chemistry, physics, energy, Earth systems, and high-performance computing, to name a few. DOE also supports national security missions such as stewardship of the nation’s nuclear stockpile. Because DOE’s mission includes so many scientific and engineering disciplines, it is only possible to provide a few examples below to illustrate how FMs might accelerate progress.

Materials Science

Materials science seeks to understand and control the relationships between structure, processing, properties, and performance across multiple spatiotemporal scales. FMs trained on experimental data, literature, and simulations offer a promising path to accelerate discovery—namely, through property prediction, retrosynthesis, and molecular generation. These models can predict properties, generate candidate structures, and guide automated experiments, reducing reliance on costly first-principles calculations (Berger 2025; Pyzer-Knapp et al. 2025). When coupled with high-throughput synthesis, they could transform ma-

terials discovery from a decades-long process into an iterative, data-driven cycle. Recent advances include MatBERT, a materials science transformer developed at Lawrence Berkeley National Laboratory (Trewartha et al. 2022), and IBM’s open-source FMs for sustainable materials design (Martineau 2024). We highlight some areas in which FMs are already being used to significant impact—namely, property prediction, retrosynthesis, and molecular generation—and also look to the future to outline areas that we believe are key to continuing to unlock value. These areas hinge on exploiting the natural multimodality and multifidelity characteristics of materials data through increasingly powerful and elegant modeling approaches. The application of FMs faces challenges such as vast chemical and structural design spaces, bridging scales, and integrating experimental, computational, and theoretical insights into predictive frameworks with quantified uncertainties (Morgan and Jacobs 2020).

Battery Technology

In battery technology, FMs are accelerating innovation from materials to management. Researchers are developing these models to rapidly screen and predict the properties of novel battery materials, such as new electrolytes, thus speeding up the discovery process (Xu et al. 2024). Furthermore, they are being used to create more sophisticated battery management systems that provide highly accurate predictions of a battery’s state of health and remaining useful life (Chan et al. 2025). By understanding the deep patterns of battery degradation, these models are helping to design safer, longer-lasting, and more efficient energy storage solutions for everything from electric vehicles to grid-scale applications.

Advanced Manufacturing

Advanced manufacturing (AM) uses computer-controlled, automated processes to produce complex components relevant to DOE’s mission. This type of manufacturing distinguishes itself from conventional mold-based or subtractive manufacturing in that it enables rapid prototyping, cost-effective experimentation, and just-in-time production of complex components as a single unit (e.g., rocket nozzles). AM is increasingly vital to DOE and its National Nuclear Security Administration (NNSA) for both energy science and national security missions, with the goal of creating parts that are “born qualified” for their intended use (Boyce 2016). Moreover, FMs offer promising solutions by integrating heterogeneous data to support tasks such as anomaly detection, process optimization, and predictive control (Autodesk 2025; Era et al. 2025; NVIDIA n.d.; Zhang et al. 2025).

Key challenges remain, as AM materials are often out of thermodynamic equilibrium, leading to undesirable properties such as low ductility or fracture toughness (Forien 2023). Developing digital twins, computational replicas of AM processes, is a major research focus across DOE and NNSA laboratories (LLNL n.d.) and an area that is being transformed by the adoption of FMs.

Weather and Earth Systems

Predicting weather and understanding Earth systems is critical to decision making (Conti 2024). Large-scale simulations of weather remain limited in the range of their time horizons and spatial resolutions due to computational constraints. FMs offer solutions by developing accurate parameterizations for subgrid-scale processes such as clouds and turbulence, identifying patterns, and improving understanding of climate dynamics. FMs for Earth systems and weather are pretrained on massive, heterogeneous Earth-system data sets and are fine-tuned for diverse downstream tasks. Data-driven weather models, such as GraphCast (Lam et al. 2023) are becoming central to several FMs that involve localized forecasts and can significantly impact applications such as agriculture and power grids. Aurora (Bodnar et al. 2025), a 1.3-billion-parameter model pretrained on more than 1 million hours of multimodal geophysical data, outperforms traditional numerical forecasts in global weather forecasting, air quality monitoring, ocean wave prediction, and tropical cyclone tracking, all at lower computational cost. Similarly, Prithvi WxC (Schmude et al. 2024), a 2.3 billion-parameter transformer model trained on 160 atmospheric variables from MERRA-2, is designed for multitask adaptation, including downscaling, extreme-event estimation, and parameterization. Projects such as ORBIT (Wang et al. 2024), a hybrid transformer model with 113 billion parameters for Earth system predictability, hold potential to accelerate climate projections, improve extreme-event forecasts, and unify disparate Earth-system modeling tasks. Key challenges include integrating real-time data assimilation, maintaining physical consistency over long prediction horizons, and scaling to capture multiscale interactions across atmosphere, ocean, land, and cryosphere.

Fusion

The quest for fusion energy involves extremely complex plasma physics and engineering challenges. FMs can accelerate the computationally demanding simulations of plasma behavior (e.g., using codes such as X-Point included Gyrokinetic Code) (Churchill 2024), help analyze the vast amounts of diagnostic data from experiments such as Doublet III D-Shaped or the International Thermonuclear Experimental Reactor (see also the agenda for the Simple Cloud-Resolving E3SM Atmosphere Model), assist in designing reactor components tolerant of extreme conditions, and potentially contribute to real-time plasma control systems needed for sustained fusion reactions. We believe that this trend will continue. In inertial confinement fusion, AI and fine-tuned FMs can help design reproducible high-fusion-gain targets. These models are pretrained on vast and diverse data sets including experimental data from tokamaks and massive simulations. They are fine-tuned for downstream tasks, such as predicting plasma disruptions, optimizing control systems in real time, improving diagnostic interpretation, and accelerating the design cycle for reactor components

(Badalassi 2023; Churchill 2024; DOE 2024). In magnetic confinement experiments, such models are being explored for real-time control to adjust magnetic fields and mitigate instabilities like tearing modes, which can severely limit performance (DOE n.d.). Beyond control, LLMs augmented with (RAG are being used to quickly access historical data and identify similar experimental conditions to guide new trials (Poore 2023). Furthermore, the design of durable fusion materials and tritium-breeding blankets that can withstand extreme reactor environments is being addressed by integrating foundation models with high-performance computing to create comprehensive simulation environments (Badalassi et al. 2023; DOE 2024; PNNL n.d.).

The committee would like to stress some of the dangers of adapting an application too quickly. In the past 2 years, several groups have sought to replace costly plasma simulations with autoregressive neural surrogates that evolve hydrodynamic and electromagnetic fields without direct partial differential equation solutions (Carey et al. 2024, 2025; Galletti et al. 2025; Gopakumar et al. 2023; Poels et al. 2023). While these are important first steps, there are fundamental challenges that must be addressed before such approaches can form the basis of a true FM for fusion, comparable to those emerging in weather forecasting. Current efforts rely heavily on Fourier neural operators (FNOs), which cannot readily accommodate the complex geometries required for magnetic confinement fusion. Moreover, autoregressive roll-outs are prone to compounding errors over long prediction horizons (McCabe et al. 2023). This issue is particularly acute in fusion, where predictions must preserve gauge symmetries and conservation laws; this is well known in conventional plasma simulation contexts within the DOE community (Sharma et al. 2020). Off-the-shelf FNOs and transformer models lack these structural guarantees. A viable FM for fusion will therefore require new approaches that ensure long-term stability and strict preservation of physical structure.

Stockpile Stewardship

The U.S. Stockpile Stewardship Program (SSP), managed by NNSA and its nuclear enterprise, aims to maintain the safety, security, and reliability of the nuclear arsenal without resuming underground testing. The national laboratories involved in these efforts have made significant progress using machine learning to obtain a deeper understanding of the relevant science and are increasingly exploring the use of FMs. These FMs are tuned with classified weapons science knowledge to gain a deeper understanding of the physics involved, thereby accelerating progress across the entire program. This represents a substantial shift toward data-driven maintenance of the stockpile.

One critical area for FM deployment is stockpile surveillance, the continuous monitoring of the health of the arsenal. FMs can be fine-tuned using a wealth of past findings and diagnostic images to rapidly assess potential deleterious

changes, helping experts quickly distinguish between changes that are not material to future performance and those that require intensive investigation via simulation and experiment. Furthermore, FMs are essential in designing digital twins that predict component failure over time—an especially difficult task. By measuring a part’s response to its dynamic environment and assimilating these data, an FM can construct a digital twin to provide advance warnings of impending failure, such as fracture due to fatigue, allowing for proactive maintenance. There is significant overlap with areas such as structural health monitoring that may be useful to adopt in this effort (see following paragraph).

Despite their potential, the use of FMs for stockpile stewardship involves significant risks and challenges. The most prominent concern is security, as classified information must be strictly controlled and only provided to staff with the necessary “need to know” clearance, a protocol that must be maintained even within secure laboratory confines. Another challenge is preventing overreliance on the guidance provided by FMs, as this could inadvertently lead to poor design decisions regarding weapon components. The DOE laboratories involved in the SSP are well aware of these issues and are actively working to mitigate these risks.

Structural Health Monitoring

FMs are gaining significant attention for structural health monitoring and infrastructure surveillance, extending their utility from high-security areas such as the nuclear SSP to civilian applications such as bridges, viaducts, and highrise buildings. FMs can absorb massive, unlabeled data sets derived from sensors—including accelerometers for vibration monitoring, imaging diagnostics, and Internet of Things devices. This generalized pretraining allows the models to learn robust, universal representations of structural behavior. Downstream tasks include anomaly detection and traffic load estimation on real-world civil infrastructure data (Benfenati et al. 2025; Bormon 2025; Hassani et al. 2024). A key application of FMs in civil infrastructure is the creation of intelligent, high-fidelity digital twins. By continuously assimilating real-time data from the physical structure (the “real twin”), FMs enable the virtual replica to accurately predict degradation, fatigue, and component failure over time. The integration of FMs into digital twins is an active area of investigation, aiming to reduce the significant manual effort typically required to create and maintain these models for cyber–physical systems (Ali et al. 2024). Although this technology promises enhanced safety and optimized resource allocation by distinguishing critical changes from nonmaterial ones, the field faces challenges related to data security, ensuring the fidelity and trustworthiness of FM-generated predictions, and managing the large computational resources required for both training and real-time inference.

Combustion

Combustion systems, including engines, gas turbines, furnaces, and scram-jets, show highly unsteady, multiscale dynamics. These dynamics stem from complex interactions among turbulence, multiphase, and reacting flows. Current physics-based simulations are too costly for extensive design or operating space exploration and cannot directly use real-world experimental data. FMs are increasingly adopted for combustion research by leveraging vast heterogeneous data sets, such as direct numerical simulations, large-eddy simulations, and experimental diagnostics, to learn universal representations of combustion phenomena (Ihme and Chung 2024). FMs can assist in the acquisition of new insights into the physics controlling flame ignition, burning rate, flame stability, and emissions in high-pressure premixed combustion of various fuels, including hydrogen. These developments are crucial for the improvement of multifidelity science-based reduced-order models, methods, and digitalization, ultimately used by U.S. industry and its clients for optimal design and operation, near-real-time risk mitigation, and maintenance. Examples of ongoing efforts include a knowledge processing framework for combustion science that integrates FMs with RAG to systematically parse literature, data sets, and simulation results, enabling automated reasoning and accelerated model development (Sharma and Raman 2024). The interfacing of combustion and machine learning is mostly focused now on adopting supervised and semi-supervised machine learning techniques to combustion problems,

Recent progress in physics-informed machine learning provides a pathway to embedding physical constraints directly into FMs, making them suitable for high-fidelity combustion simulations (Cao et al. 2026). The adoption of an inverse modeling approach (Karnakov et al. 2024) and the extension of these efforts in order to account for proper validation and verification (McGreivy and Hakim 2024) within an FM framework holds great potential for combustion science, an area central to the mission of DOE.

National Security

In addition to the potential benefits described above, FMs can bolster other national security missions where DOE plays an important role:

• Nonproliferation and threat detection. FMs can process large, heterogeneous data sets (e.g., satellite imagery, sensor data) to identify nuclear proliferation activities or emerging threats.
• Strategic analysis. They can assist analysts by synthesizing information from technical, geopolitical, and open-source materials to support strategic decision making.

FMs offer powerful tools for managing and securing energy infrastructure, such as the following:

• Grid management and optimization. FMs trained on operational data, weather patterns, and energy markets can enhance load forecasting, predict renewable generation (solar, wind), and optimize grid operations for efficiency and stability.
• Resilience and threat mitigation. By analyzing complex system inter-dependencies, FMs can identify vulnerabilities to physical threats (e.g., extreme weather) or cyberattacks. They can also assist in developing response and recovery strategies, complementing planning tools such as the North American Energy Resilience Model. The concept of “GridFMs”—FMs trained on diverse grid data—could significantly advance predictive capabilities, especially for cascading failure scenarios.

Although offering important benefits, FMs also pose risks if misused. The adversarial use of FMs, particularly LLMs, presents significant security risks that can be broadly summarized in two categories: attacks targeting the model itself and attacks leveraging the model as a weapon.

Attacks against the model exploit its vulnerabilities to subvert its intended function or extract sensitive data. This includes prompt injection (or “jailbreaking”), where an attacker crafts input to bypass safety filters and force the model to generate harmful or restricted content. Another major threat is data poisoning, which occurs when malicious data are subtly inserted into the training set, creating hidden backdoors or permanently degrading the model’s accuracy. Finally, risks such as model inversion and model stealing compromise confidentiality by allowing adversaries to reconstruct sensitive training data or illegally copy the model’s proprietary intelligence.

The second major risk involves using powerful FMs to accelerate and scale traditional cyberattacks. Adversaries leverage these tools to generate highly convincing and personalized phishing e-mails and synthetic media (deepfakes), vastly increasing the success rate of social engineering. FMs also lower the barrier for technical attacks by helping actors write and optimize malicious code or rapidly identify software vulnerabilities, making advanced cyberthreats more common. Furthermore, the complexity of integrating these models into larger systems creates new supply chain risks. For example, a successful prompt injection against an LLM that is integrated with an external tool (i.e., a database) can be used to execute a traditional command injection attack against the connected system, demonstrating that the AI model itself can become a single point of failure and a gateway to broader network compromise.

Users of FMs should invest in AI assurance, red teaming, and development of countermeasures against adversarial applications of FMs, aligning with strategies such as Advance Simulation and Computing’s Artificial Intelligence

for Nuclear Deterrence program and the Frontiers in Artificial Intelligence for Science, Security and Technology’s trustworthy AI pillar.

Recommendation 3-3: To address potential security risks arising from the adversarial use of foundation models, the Department of Energy should explore strategies for artificial intelligence assurance, red teaming, and development of countermeasures.

REFERENCES