DESC0025068

Artificial intelligence (AI) systems built on large multimodal foundation models have demonstrated impressive capabilities, but also exhibit vulnerabilities such as bias, hallucinations, lack of transparency, and potential for malicious misuse. Thoroughly evaluating the trustworthiness, safety, and security of these AI systems is resource intensive yet crucial for responsible development and deployment, especially in high-stakes domains like national security. There is a critical need for more efficient, comprehensive, and automated testing platforms to identify and mitigate risks in multimodal AI systems, to ensure their robust and reliable operation while protecting against unintended consequences. The proposed effort aims to address the challenge of evaluating and securing multimodal AI systems by developing ART-FM, a novel automated experimentation testbed for red teaming and evaluation of static and dynamic AI systems built upon multimodal foundation models. ART-FM will provide a cloud-based sandbox environment for systematic probing and testing of AI systems across their entire lifecycle, from data to deployment, enabling the proactive discovery of vulnerabilities, risks, and unintended behaviors. The project approach combines Aptima's expertise in model-agnostic evaluation, causal model explanation, and interactive test analytics with Exostellar's transparent cloud computing and resource optimization technologies. This will yield an integrated platform for AI researchers, developersa, nd auditors to efficiently assess and harden multimodal AI systems against a wide range of integrity, reliability, and security threats, thereby accelerating the development of safe and trustworthy AI for national security and beyond. In Phase I, Aptima and Exostellar will develop the ART-FM experimentation framework to effectively evaluate and red team multimodal foundation models (MFMs). This will involve designing the cloud sandbox architecture, developing reasoning analytics modules for vulnerability detection and explanation, implementing automated testing workflows, and creating a unified scoring framework for AI trustworthiness. The team will also investigate the performance and mobility of the ART-FM framework across different cloud environments while conducting initial evaluations on state-of-the-art MFMs. Finally, a commercialization and transition plan will be developed to guide the future development and deployment of ART-FM. If ART-FM progresses to Phase II and beyond, it has the potential to revolutionize the development and deployment of trustworthy AI systems across a wide range of applications. Beyond national security use cases, ART-FM could be leveraged by commercial enterprises to ensure the safety, reliability, and fairness of AI systems in domains such as healthcare and energy. In the public sector, ART-FM could support the responsible adoption of AI by government agencies, enabling them to deliver more effective and efficient services to citizens while mitigating risks.