AI Leaders Propose SAFE Guidelines for Cybersecurity Transparency

Open Secure AI Alliance is developing Shared AI Findings Exchange (SAFE) guidelines to strengthen agentic AI security, with participation from over 120 organizations. The Linux Foundation has begun soliciting feedback on the SAFE draft in conjunction with its annual Black Hat event.

SAFE proposes a system to confidentially collect and analyze AI security incidents and near-misses, notify affected organizations, and publicly release evidence-based operational recommendations by identifying recurring control failures. This transforms lessons learned from individual incidents into collective defense capabilities for the entire AI ecosystem.

NVIDIA, Cisco, CrowdStrike, Hugging Face, Red Hat, and others are participating in drafting the initial proposal alongside the Linux Foundation. The alliance views agents not merely as models but as systems composed of identity controls, harnesses, guardrails, logs, and evaluation systems, addressing the entire security stack.

Key security tools and models released by NVIDIA include:

• NOOA: A research harness for testing, tracing, auditing, and managing agent behavior
• NVIDIA OpenShell: A runtime that restricts the scope of what agents can access or execute
• NVIDIA Nemotron, Cosmos, Isaac GR00T, BioNeMo, Alpamayo: An open model family for agentic AI, physical AI, robotics, healthcare, and autonomous driving
• Verified Agent Skills: A system for managing agent capabilities that checks for prompt injection and tool poisoning risks and provides cryptographic signatures and documentation
• NeMo Guardrails, NeMo Anonymizer, NeMo Safe Synthesizer: Tools for enforcing safety policies, protecting sensitive data, and generating privacy-preserving synthetic data
• Garak: An open-source LLM vulnerability scanner that checks for data exfiltration, prompt injection, and jailbreak scenarios