NVIDIA and more than 35 technology, cybersecurity, cloud computing and artificial intelligence organizations have formed the Open Secure AI Alliance to develop open technologies for protecting software, AI agents and critical digital infrastructure.
The initiative builds on work conducted through the Linux Foundation’s Akrites project and the Open Source Security Foundation community. Its members plan to collaborate on open models, agent harnesses, security tools, evaluation systems and vulnerability remediation methods.
Inaugural partners include NVIDIA, Adobe, Cadence, Capital One, Cisco, Cloudera, Cloudflare, Cognition, CrowdStrike, Databricks, Dell Technologies, DoorDash, Elastic, HPE, Hugging Face, IBM, LangChain, the Linux Foundation, Microsoft, NAVER, NetApp, Nous Research, OpenClaw, Palantir, Palo Alto Networks, Red Hat, Reflection AI, Salesforce, SAP, ServiceNow, Siemens, SK Telecom, Snowflake, SpacexAI, Synopsys, Thinking Machines Lab and TrendAI.
The alliance is based on the view that cybersecurity teams need access to both closed and open AI systems. Open models can be inspected, customized and operated on privately controlled infrastructure, potentially helping organizations respond to threats without sending sensitive information to an outside provider.
NVIDIA said open systems can also reduce dependence on a single vendor and enable companies, governments and researchers to build security tools across a broader technology ecosystem.
The alliance acknowledged that open models can be modified or misused, including for cyberattacks. However, its members argue that advanced closed systems can also be abused and that restricting model access does not eliminate malicious activity.
The group is advocating for a security framework that combines open development with safeguards, usage policies, testing, evaluation and rapid vulnerability remediation.
The alliance cited a recent security incident involving Hugging Face as an example of why defenders may need models that they can run and adapt themselves. According to NVIDIA, closed AI tools blocked portions of the forensic analysis because they could not reliably distinguish defensive investigation from malicious activity.
Hugging Face instead operated the open-weight GLM 5.2 model on its own infrastructure to analyze more than 17,000 actions and assist with containing the intrusion. NVIDIA said the incident demonstrated how restrictions imposed by external AI services can limit cybersecurity teams during time-sensitive investigations.
A central focus of the alliance will be the complete AI agent stack rather than only the underlying language model. An AI agent may include identity systems, permissions, execution environments, memory, logs, guardrails, evaluation tools and software that determines how the model interacts with other systems.
Open agent harnesses can allow defenders to examine how these components work together, test their behavior and identify security weaknesses before deployment.
NVIDIA is contributing models, model weights, data and agent harness research to the alliance. The company also introduced the open-source NVIDIA Labs Object-Oriented Agent project, known as NOOA.
NOOA is a research framework designed to help developers integrate models with agent harnesses while making agent behavior easier to test, trace, audit and govern. This could help organizations understand which actions an agent took, what information it used and whether its behavior remained within defined permissions.
Other alliance members are contributing technologies across identity, software supply chain security, model storage, vulnerability scanning and coding.
HPE contributes to SPIFFE and SPIRE, open standards and software for assigning cryptographically verifiable identities to workloads. These systems can help ensure that only authorized AI agents and services communicate with enterprise applications or access protected resources.
Hugging Face has offered Safetensors to the PyTorch Foundation. Safetensors is a model-weight storage format designed to improve transparency and prevent arbitrary remote code execution when model files are loaded.
IBM and Red Hat are contributing Lightwell, which extends security across the open-source software supply chain through digitally signed patches.
Microsoft’s MDASH system uses multiple specialized AI agents to identify, analyze and validate exploitable software vulnerabilities.
SpacexAI has open-sourced Grok Build, a terminal-based AI coding agent, and plans to make weights from the Grok model family available to developers and researchers.
The alliance is also calling on governments and regulators to recognize open AI models, harnesses and security tools as defensive assets. Its members argue that broad restrictions on open frontier systems could concentrate cybersecurity capabilities among a small number of closed providers.
Instead, the organization is encouraging investment in shared datasets, attack simulators, evaluation frameworks, red-teaming systems and other open infrastructure for AI defense.
The Open Secure AI Alliance plans to provide a forum through which companies, researchers and public-sector organizations can jointly develop tools for securing increasingly autonomous AI systems.
