The Open Secure AI Alliance, comprising over 120 organizations, is launching SAFE — a framework for sharing AI cybersecu
The Open Secure AI Alliance announced new guidelines during Black Hat conference to address the emerging cybersecurity challenges posed by agentic AI systems. The Linux Foundation unveiled a Request for Comments on Shared AI Findings Exchange (SAFE), a proposed set of guidelines designed to transform agentic cybersecurity incidents into shared protection across the entire ecosystem. NVIDIA, Cisco, CrowdStrike, Hugging Face and Red Hat are among the founding contributors working with the Linux Foundation on this initiative.
The SAFE guidelines propose to confidentially collect and analyze AI incidents and near misses, inform impacted parties, identify recurring control failures and publish evidence-based operating recommendations that reduce systemic risk. The framework recognizes that securing AI agents requires more than vulnerability scanning — agents are complete systems encompassing identity controls, harnesses, guardrails, logs and evaluation mechanisms.
Alliance members are contributing tools and technologies across multiple layers of the AI security stack. NVIDIA is releasing the Object-Oriented Agent (NOOA) research harness to make agent behavior easier to test, trace, audit and govern, along with OpenShell runtime to enforce security and privacy controls at the agent level. The company is also contributing open model families including Nemotron for agentic AI, Cosmos for physical AI, Isaac GR00T for robotics, and BioNeMo for healthcare applications, all shipped with open weights and training techniques. NVIDIA's open source verified agent skills are cataloged, scanned for risks such as prompt injection and tools poisoning, cryptographically signed and documented with skill cards. Additional tools include NeMo Guardrails for safety policies, NeMo Anonymizer for data protection and Garak, an open source LLM vulnerability scanner.
Other contributors include Okta, developing reference implementations for agent identity and access; Palo Alto Networks, contributing Agent Guard and Agent Watch from its Idira platform; and Red Hat, launching asago to map governance requirements from NIST, OWASP and the EU AI Act directly to runtime controls. Amazon, which joined the alliance today, contributes Strands Agents for building AI agents with full visibility and Cedar, an open source authorization language for enforcing fine-grained access controls. Microsoft AI Red Team has released PyRIT for automated red teaming, RAMPART for converting findings into repeatable tests and Assert for converting safety requirements into executable evaluations. Capital One open sourced VulnHunter for agentic AI code security, while Cloudflare offers its Vulnerability Discovery Harness and Wiz contributes Atlas, an autonomous vulnerability research engine. Cisco released DefenseClaw as an agentic governance layer and two Antares security small language models for vulnerability pinpointing. CrowdStrike fine-tuned NVIDIA Nemotron Nano for cyber defense, achieving 96% accuracy in generating investigation queries within Falcon LogScale.
The alliance emphasizes that the strongest security emerges from layered defenses and community collaboration, with defenders learning from each other openly and at speed to collectively address threats to infrastructure and intellectual property.