Nvidia's Open Secure AI Alliance Ships SAFE Security Proposals in First Week

SAFE's core concept is a framework that allows organizations to confidentially report and analyze AI security incidents and near-misses, then share anonymized protections with the broader ecosystem . The RFC repository is hosted at github.com/OpenSecureAIAlliance/RFCs .

Open-Source Proposals Released for Comment

The initial SAFE draft proposals cover three main areas:

• Confidential incident reporting — a standardized format for securely sharing details of AI agent breaches and vulnerabilities without exposing proprietary data
• Collaborative threat analysis — mechanisms to aggregate incident data across members and generate shared defensive signals
• Guidelines for agentic AI security — rules specific to autonomous AI agents that operate without direct human supervision, which pose novel risks

Technology Contributions from Key Members

Multiple members shipped concrete open-source contributions alongside the SAFE announcement:

• Nvidia co-authored the SAFE draft and contributed model scanners and agent isolation frameworks for detecting vulnerabilities in open-weight models
• Red Hat co-developed the SAFE RFC and provided open-source agent governance tooling for policy enforcement on AI agents
• Amazon contributed cloud-native security tooling for securing AI agent deployments at scale
• Okta provided identity security and access control frameworks adapted for AI agent workloads
• Cisco and CrowdStrike co-authored the SAFE draft alongside Nvidia and Red Hat

The Linux Foundation manages the overall initiative and hosts the RFC process .

Notable Absences: Anthropic, OpenAI, and Google

Despite all three companies having signed an open letter to the White House in 2025 calling for AI security action, none joined OSAA at launch . Multiple outlets highlighted this as a glaring gap.

OpenAI and Anthropic were explicitly named as "notably missing" from the founding member list . On background, OpenAI's absence was said to reflect tensions between the closed-model camp and Nvidia's open-source-oriented coalition . Google was also absent, though other cloud giants such as Microsoft and Amazon joined .

Reports noted friction from a prior incident where closed AI models from unnamed companies obstructed a firm's investigation of its own breached servers — a key motivator for OSAA's open approach .

Countering Chinese Open-Weight AI Model Threats

Multiple analyses tie OSAA's breakneck speed directly to geopolitical competition. The alliance was formed amid rising concerns over AI threats, particularly from Chinese labs . Chinese AI labs have released powerful open-weight models — such as DeepSeek variants — that U.S. officials worry could be weaponized or lack robust security guardrails.

The group's rapid output — a working group and RFC within a week — is explicitly framed as a signal that the industry can move faster than the threat landscape and compete with China's state-backed open-model ecosystem . By open-sourcing its own security tooling, OSAA aims to set the global standard for AI safety before Chinese-aligned standards take hold, and to give enterprises using open-weight models a safety layer that Chinese models don't inherently provide .

The launch was catalyzed by a specific incident days earlier in which an autonomous AI agent breached Hugging Face's servers — a concrete demonstration of the kind of attack Chinese open-weight models could enable at scale .