AI Leaders Propose SAFE Guidelines for Cybersecurity Transparency

The Open Secure AI Alliance, comprised of over 120 organizations, is drafting a set of guidelines called SAFE (Shared AI Findings Exchange) to enhance cybersecurity for agentic AI systems. This initiative, being worked on by a working group, includes contributions from prominent tech companies such as NVIDIA, Cisco, CrowdStrike, Hugging Face, and Red Hat.

The goal of the SAFE guidelines is to develop a system for confidentially collecting, analyzing, and sharing information about AI incidents and near misses, notifying those affected, identifying recurring control failures, and publishing evidence-based operating recommendations. The alliance emphasizes the importance of cybersecurity in the rapidly evolving technology landscape, noting that each major technological shift creates new attack surfaces.

They argue that rapid response times are crucial for protecting infrastructure and intellectual property, and that open sharing of threat intelligence among trusted ecosystems can significantly boost collective defense. The SAFE framework is part of the alliance's broader commitment to creating and sharing open, inspectable tools across the entire AI security stack.

This includes contributions from NVIDIA, which offers open-source cybersecurity software and models running from the lowest layers of the stack, such as the NVIDIA Labs Object-Oriented Agent (NOOA) research harness, to high-level models like NVIDIA Alpamayo, the world's largest model for autonomous vehicles. The alliance members are also developing tools for identity and permissions, harnesses, runtime guardrails, security AI models, observability, evaluation, data security and privacy, availability and resilience, and more.

Examples include Okta's work on agent identity and access, Palo Alto Networks' Idira tools for identity security, and Red Hat's asago project that aligns governance requirements with agent activities.