AIExploit: automated detection of exploits in AI/ML models for red team penetration testing

The AIExploit platform is a pioneering automated tool designed to enhance the U.S. Army's operational readiness by detecting and preventing unforeseen risks in AI models. This innovation allows the Army to deploy AI in contested environments, empowering personnel to team effectively with AI, thereby gaining a technological edge over adversaries. Our platform introduces a revolutionary capability: on-demand, automated, no-code AI red teaming. This is a significant shift from traditional operational testing, which typically requires the presence of AI/ML developers. Our solution enables real-time vulnerability assessments in the field, providing immediate insights to commanders without developer intervention. This transformative approach enhances the U.S. Army's operational landscape in human-AI teaming, accelerating the closure of the kill chain more efficiently than adversaries. The motivation behind AIExploit stems from the need for operationalizing emerging technologies like AI, ML, and autonomy, which are crucial for the U.S. National Defense Strategy. These technologies offer competitive advantages but also introduce new security risks. Military leadership faces a dilemma: either adopt new technologies with inherent security vulnerabilities or delay their adoption until a full understanding of the risks is achieved, potentially falling behind adversaries. The AIExploit platform addresses this challenge by enabling operators to understand AI's strengths and limitations better, thus making more informed decisions while leveraging AI capabilities. Yrikka's solution builds on the success of its Phase I feasibility study, which resulted in the development of a TRL 4 prototype automated red-teaming tool for context-based AI Risk Management. This tool will enable Army operators and commanders to efficiently identify exploits and vulnerabilities in AI models, crucial for critical decision-making tasks. Our goal in this Phase II project is to evolve the current TRL 4 prototype of AIExploit into a TRL 7 software module. This module will automate the detection of potential exploits in computer vision AI/ML models used by the U.S. Army, supporting modern armament systems and the development of Advanced Fire Control Technology (AFCT).