The Open Secure AI Alliance, a coalition of more than 120 organizations including Cisco, CrowdStrike, Hugging Face, NVIDIA, Red Hat, Amazon, and Microsoft, published a Request for Comments for the Shared AI Findings Exchange (SAFE) Working Group on August 4, 2026, at the Black Hat conference in Las Vegas. The Linux Foundation issued the RFC, which proposes a vendor-neutral framework for confidentially reporting and analyzing agentic AI security incidents and near misses, then converting findings into reusable defenses. SAFE is modeled after NASA's Aviation Safety Reporting System for blameless reporting of AI operational failures. It is not a finished standard, reporting authority, or completed specification; it remains an open proposal seeking community contributions on GitHub.
The framework defines a notification ladder: affected organizations must be notified immediately, customers with demonstrable exposure within 72 hours, a confidential initial report within four business days, and a preliminary public report within 30 days. Members are required to report near misses, not only confirmed incidents. SAFE covers model, tools, permissions, runtime environment, safeguards, monitoring, and human oversight when reviewing incidents, and intends to publish reusable tests, detection rules, and incident response guidance based on collected data.
Alongside the RFC, alliance members released open-source tools. NVIDIA contributed the NOOA research harness, OpenShell runtime sandbox, verified agent skills with cryptographic signing, NeMo Guardrails, and the Garak LLM vulnerability scanner. CrowdStrike fine-tuned NVIDIA Nemotron Nano for cyber defense, reporting 96% accuracy in generating investigation queries and outperforming larger models on Security Operations Center detection triage. Uber open-sourced ADR (Agentic AI Detection and Response), a production system processing over 200,000 agent sessions per day across 30,000 endpoints that reconstructs the full causal chain of AI agent activity. New members Amazon and Visa contributed Strands Agents, Cedar authorization language, and a Vulnerability Agentic Harness.