SAFE-AI Framework for Securing AI-Enabled Systems

This technical report presents the SAFE-AI framework, which focuses on securing AI-enabled systems by evaluating the risks introduced by AI technologies. The document outlines the importance of selecting appropriate security controls corresponding to the identified risks, emphasizing the need to address AI-specific threats systematically. The SAFE-AI framework is based on guidelines set by the National Institute of Standards and Technology (NIST) and integrates with the MITRE Adversarial Threat Landscape for Artificial Intelligence Systems (ATLAS) framework. The report documents various adversarial threats, including adversarial inputs, model biases, and supply chain vulnerabilities, which can compromise AI systems. It also discusses the high stakes involved in using third-party libraries and pre-trained models that may harbor hidden vulnerabilities. The report recommends utilizing controls from the NIST control catalog, Special Publication 800-53, Rev 5, and offers mitigations along with a discussion on residual risks associated with AI technologies. The intended audience includes cybersecurity professionals responsible for securing information systems and developing system security plans.