Linux Foundation proposes SAFE guidelines for agentic AI security

The Linux Foundation shared a Request for Comments for Shared AI Findings Exchange, or SAFE, as members of the Open Secure AI Alliance work on guidelines for agentic AI cybersecurity. The alliance, now more than 120 organizations strong, is developing the proposal with contributors including NVIDIA, Cisco, CrowdStrike, Hugging Face and Red Hat.

SAFE is designed to turn AI incidents and near misses into shared protections across the ecosystem. The proposed guidelines call for confidential collection and analysis of incidents, notification of impacted parties, identification of recurring control failures and publication of evidence-based operating recommendations to reduce systemic risk.

The effort sits alongside open contributions across the AI security stack. NVIDIA highlighted tools including the NOOA research harness, OpenShell runtime, open model families, verified agent skills, NeMo safety tools and Garak, its open source LLM vulnerability scanner. Other alliance members are contributing identity controls, authorization tools, agent harnesses, red-team tooling, safety models, observability systems and resilience features.

Several projects target production defenses for agentic systems. CrowdStrike reported internal testing with 96% accuracy for generating investigation queries in Falcon LogScale, while Uber said ADR supports more than 200,000 agent sessions per day across 30,000 endpoints. LangChain and Veeam are adding recovery and resilience capabilities for AI workloads and supporting infrastructure.