The Open Secure AI Alliance, now 120+ organizations strong, has proposed the SAFE (Shared AI Findings Exchange) guidelines at Black Hat 2024 to improve cybersecurity transparency for agentic AI systems. The framework aims to confidentially collect and analyze AI incidents, identify recurring control failures, and publish evidence-based recommendations. NVIDIA and alliance members including Cisco, CrowdStrike, Amazon, Microsoft, and others are contributing open source tools across the full AI security stack — covering agent identity and permissions, harnesses, runtime guardrails, specialized security models, observability, and resilience. Notable contributions include NVIDIA's NOOA research harness, OpenShell runtime, Garak LLM vulnerability scanner, Microsoft's PyRIT and RAMPART, CrowdStrike's fine-tuned Nemotron Nano for cyber defense, and LangChain's resilience capabilities for agent recovery.
