15 Over-optimization in some functions of the military presents noteworthy risks to resilience, while other areas stand to become more resilient, such as improvements in monitoring of new and emerging data feeds for predictive maintenance. Branches of the military currently encounter significant cost and scheduling challenges due to maintenance issues with systems and equipment. While historically, the military’s approach has been preventative in nature—to create standardized scheduling for maintenance reviews and address problems as they arise—Generative AI and broader AI-family technologies stand to move the military’s approach to maintenance toward one that is predictive in nature. In the Navy, for example, Material Readiness Cards are used to state when maintenance should be expected; however, the system involves entry into static intake forms involving manual reporting, without proactive alert systems. David Miller, General Manager for Intellisense’s fast-growing Sensors and Integrated Systems segment, further notes that the key to enabling reasoning models to support predictive maintenance is the sensors themselves: Past estimates indicate that implementing sensors, and processing their data streams, could lead to better maintenance cycles that save branches like the Air Force as much as $15 billion annually.23 Overall, the clear tension between dependency and efficiency is a theme that the military will consistently need to balance as it increasingly integrates AI, and particularly Generative AI tools. Teams may find that in certain areas where enabling technology and infrastructure can be modified to support Generative AI, the tech stack should not always be modified because of potential drawbacks to resiliency. Weighing tradeoffs between technological dependency and gained efficiency is critical. To identify these tradeoffs, we recommend following established best-practices in Red-Teaming as soon as teams finish the evaluation phase of a Gen-AI integration (described below) and thereby understand how Gen-AI tools will impact their processes. Red-Teaming allows teams integrating Gen-AI tools to assess the impact on the organization if an adversarial attack disabled, eliminated, or corrupted the capabilities of a given tool. Examples could include Direct Disruption of Service (DDoS) attacks that interfere with communication channels relaying critical data to models; hacking efforts, including trojans or other malware, that seek to covertly skew or corrupt data at its real-time source feed. Siloed and infrequently monitored data repositories are a particular area of vulnerability, as they can impact Gen-AI outputs when later integrated into models using RAG. Red-Teaming can also pinpoint vulnerabilities to strategies developed by adversarial AI that directly target the DOD’s own AI tools and enabling assets. Each scenario must be evaluated through several lenses, including full-automation (total dependency on Gen-AI tools), collaboration (semi-dependency on Gen-AI tools), and abstention (no dependency on Gen-AI tools). Also, tracing the impact throughout enterprise or mission activities can help establish adversarial resilience that might otherwise be eroded if critical functions of the DOD become over-dependent on Generative AI. We expect that collaborative Gen-AI integrations will present the safest and most resilient implementations of the tools in the near term, including ‘humans in the loop’ at a given tool’s data input and output stages to identify and mitigate suspicious trends or potential threats. Overall, however, before contracting a vendor, we recommend that Red- Teaming analysis is conducted to determine the tool’s impact on resiliency for the team it serves. One example of predictive maintenance technologies is embedded sensors in systems that monitor vibrations in vehicle transmissions. As the gears wear out, they become noisier, indicating maintenance needs, allowing the Army to provide maintenance timelier. “ Generative AI Adoption in the US Military Overall, the clear tension between dependency and efficiency is a theme that the military will consistently need to balance as it increasingly integrates AI, and particularly Generative AI tools.