A featured contribution from Leadership Perspectives, a curated forum for enterprise technology leaders, nominated by our subscribers and vetted by the CIOApplications Editorial Board.
Lockheed Martin
How we Ensure Responsible AI at Lockheed Martin
Generative Artificial Intelligence (GAI) promises to enhance nearly all aspects of life from learning to creating to living (Stanford 2025) with forecasts of adding trillions of dollars annually to the global economy (McKinsey 2023, WEF 2024). Yet fears of job loss or uncontrollable agents threaten full adoption. In the former case, past technology revolutions in cars, computers and robots have each proven that while some job disruption inevitably occurs with transformative technology, cost reductions and new supply creation drive very significant job creation. Indeed, already we are observing 10-30 percent improvements in productivity in use cases of Large Language Models (LLMs) occurring across engineering, manufacturing and customer service, with adopting companies experiencing growth in their more broadly available products and services which are simultaneously benefiting from reductions in the cost of using LLMs.
At Lockheed Martin, we are advancing AI and machine learning across hundreds of use cases in areas including modelbased design, software development, customer service, proposal generation, and intelligence (Maybury, Forrest and O’Donnell, 2025). .
We have accelerated this progress and address concerns of uncontrollable agents by implementing strong policies and training on responsible AI and by adopting a centralized, open and secure AI Factory which we recently made available commercially to help secure suppliers and regulated industries (AstrisAI 2025)
By collaborating closely with major LLM providers such as Google, Meta, Microsoft, OpenAI, NVDIA, MistralAI and others, we ensure the latest and most resilient models are available to developers and end users providing clarity on their capabilities and responsible application.
Responsible AI includes understanding the limits of current technology and applying appropriate guardrails to mitigate undesirable consequences. For example, today we recognize that LLMs need to have controls to counter any vulnerabilities to their confidentiality, integrity and availability to ensure safety, privacy and security for all which we have incorporated into our own Lockheed Martin AI Factory and AstrisAI (a Lockheed Martin subsidiary). For example, to undermine confidentiality, attackers can create methods to leak privacy or security information. They can undermine integrity by poisoning training data or exploiting inference weaknesses. Finally, they can deny availability by holding training data or models for ransom. Just as adversaries can attack the large language models directly, adversaries can employ LLMs as tools against victims. For example, they can employ LLMs to learn vulnerabilities at speed and scale, corrupt integrity through high-fidelity but erroneous synthetic training data generation and deny availability by flooding LLMs with high fidelity human-like access. The good news is that LLM developers and defenders such as Lockheed Martin, AstrisAI and our collaborators are working to identify the root causes and associated harms of biased, brittle and baroque LLMs to create countermeasures that will enhance their resilience (Maybury 2025). At Lockheed Martin and AstrisAI we increase resilience by diversifying training data and teams, employing red teams, applying prompt guardrails, using knowledge graphs to validate facts, employing certainty management and providing end users explanations of responses.
The potential value of generative and agentic AI is very high, but progress must be governed with responsible engineering. We are committed to improving the correctness, coherence and clarity of LLMs to ensure safety and security. Making these systems more resilient and more available to the broader defense and commercial community through organizations such as AstrisAI, we intend for all to benefit safely from this transformative technology.
We are advancing AI and machine learning across hundreds of use cases in areas including model-based design, software development, customer service, proposal generation,and intelligence
By collaborating closely with major LLM providers such as Google, Meta, Microsoft, OpenAI, NVDIA, MistralAI and others, we ensure the latest and most resilient models are available to developers and end users providing clarity on their capabilities and responsible application.
Responsible AI includes understanding the limits of current technology and applying appropriate guardrails to mitigate undesirable consequences. For example, today we recognize that LLMs need to have controls to counter any vulnerabilities to their confidentiality, integrity and availability to ensure safety, privacy and security for all which we have incorporated into our own Lockheed Martin AI Factory and AstrisAI (a Lockheed Martin subsidiary). For example, to undermine confidentiality, attackers can create methods to leak privacy or security information. They can undermine integrity by poisoning training data or exploiting inference weaknesses. Finally, they can deny availability by holding training data or models for ransom. Just as adversaries can attack the large language models directly, adversaries can employ LLMs as tools against victims. For example, they can employ LLMs to learn vulnerabilities at speed and scale, corrupt integrity through high-fidelity but erroneous synthetic training data generation and deny availability by flooding LLMs with high fidelity human-like access. The good news is that LLM developers and defenders such as Lockheed Martin, AstrisAI and our collaborators are working to identify the root causes and associated harms of biased, brittle and baroque LLMs to create countermeasures that will enhance their resilience (Maybury 2025). At Lockheed Martin and AstrisAI we increase resilience by diversifying training data and teams, employing red teams, applying prompt guardrails, using knowledge graphs to validate facts, employing certainty management and providing end users explanations of responses.
The potential value of generative and agentic AI is very high, but progress must be governed with responsible engineering. We are committed to improving the correctness, coherence and clarity of LLMs to ensure safety and security. Making these systems more resilient and more available to the broader defense and commercial community through organizations such as AstrisAI, we intend for all to benefit safely from this transformative technology.
The articles from these contributors are based on their personal expertise and viewpoints, and do not necessarily reflect the opinions of their employers or affiliated organizations.
