Using LLMs to streamline the proposal review process

This proposal targets the Air Force Materiel Command (AFMC) Line of Effort #3 to Revolutionize Our Processes. Adoption of AI technology has been deemed necessary to modernize our digital systems and maintain a competitive edge over adversaries, but this must be paired with proper security frameworks that acknowledge the risks inherent to these AI systems. It is only a matter of time before adversaries recognize those risks and utilize them to take advantage of this rapidly growing technology. Contracting officers (COs) in the USAF have expressed interest in large language model (LLM) products that propose to expediate the contract review process for the purpose of strengthening DoD readiness levels. Online LLMs are uniquely susceptible to an attack referred to as prompt injection, where user or input data can change the behavior and output of LLMs. This can lead to the spread of misinformation or leaking of secure information. Our primary focus in this project is on bolstering the security of our LLM commercial off the shelf product, CLAWS, with reinforcement learning AI that can detect and mitigate threats specific to LLMs. A reinforcement learning approach will ensure that the planned cyber security tool will be adaptive in the face of evolving threats. Our key goal in ensuring safe, responsible AI aligns with the mission of the Deputy Secretary of DefenseÆs Generative AI and LLM Task Force, Task Force Lima, for the exploration of responsible generative AI capabilities.