NVIDIA and over 100 partners launched the Open Secure AI Alliance to develop open AI models, harnesses and tools that cy

Open source software underpins global critical infrastructure across cloud computing, financial services, manufacturing, telecommunications, government and internet services. Cybersecurity ranks among the top three beneficiaries of open source, and as the world moves into AI-driven defense, the United States and its partners face a fundamental choice: whether the systems protecting critical infrastructure will remain locked inside proprietary platforms or be built on open models, harnesses and tools that defenders can study, adapt and deploy.

While the world needs both closed and open models, open models are essential for cybersecurity because they democratize defensive capabilities, increase transparency for defenders, enable cyber defense while protecting data, and provide customizable, localized controls that complement frontier closed models. Open source enables massively distributed, community-driven defense with no single point of failure. Open models carry risks of misuse like any powerful technology, but those risks are not unique to open systems and must be managed wherever advanced AI is deployed.

The recent Hugging Face security incident illustrated this principle. When the company's closed AI tools could not distinguish attackers from defenders and blocked essential forensic analysis, Hugging Face deployed the open-weight GLM 5.2 model on its own infrastructure to analyze more than 17,000 actions and contain the intrusion. The incident demonstrated a critical truth: when defenders cannot inspect, adapt and run advanced AI on their own infrastructure, their response capability is constrained at the moment speed matters most.

The Open Secure AI Alliance brings together over 100 leaders from cloud computing, cybersecurity, enterprise software, open source foundations and AI research—including NVIDIA, Adobe, Amdocs, AHEAD, Aible, Akamai, Amazon, Anyscale, Arcade Dev, Arteris, Atlassian, Box, Cadence, Canonical, Capital One, Check Point, Checkmarx, Cisco, ClickHouse, Cloudera, Cloudflare, Cloudsmith, Cognition, Cohere, Cohesity, ControlPlane, Commvault, CrowdStrike, Crusoe, Cyberhaven, Databricks, Datadog, Dataiku, DDN, Dell Technologies, DepthFirst AI, Docker, DoorDash, Dream Security, Echo, Elastic, EleutherAI, Endor Labs, Exaforce, F5, Factory AI, Fireworks AI, Fortanix, Fortinet, G42, Genspark, GitHub, Glean, H2O.ai, HPE, Hugging Face, IBM, Infoblox, Infosys, Intel, JFrog, Kindo AI, Kong, Kyndryl, LangChain, Ledger, Lenovo, the Linux Foundation, Microsoft, Mirantis, Mistral, Mozilla, NAVER, NEAR AI, NetApp, Netskope, Nokia, Nous Research, NTT, Nutanix, Okta, OpenClaw, OpenHands, Palantir, Palo Alto Networks, Perplexity, Pinterest, Poolside, Red Hat, Reflection AI, Replit, Rubrik, Salesforce, Samsung Electronics, SAP, SentinelOne, ServiceNow, Siemens, SK Telecom, Snowflake, Snyk, Sonar, SpaceXAI, Spectro Cloud, SUSE, Synopsys, Thales, Thinking Machines Lab, TrendAI, Uber, Upwind, UiPath, Veeam, Visa, vLLM, VMware by Broadcom, VotalAI, Wiz, Workday, World Wide Technology, Zenity and Zscaler.

Some argue that open models are inherently unsafe because they can be misused for cyberattacks or modified to remove guardrails. While these risks are real, they do not disappear in closed systems, and keeping model weights closed does not prevent determined attackers from accessing or exploiting powerful AI. The appropriate response is to pair openness with strong safeguards, clear rules against malicious misuse, rigorous evaluation and rapid remediation. In cybersecurity, the safer path is the one that gives more defenders the ability to test, verify and strengthen critical systems.

An AI agent is not simply a language model but a complex system built from models, harnesses and guardrails. Real AI safety depends on the full agent stack—identity, permissions, harnesses, guardrails, logs and evaluation—not merely whether weights are open or closed. Open harnesses and tools make these controls easier for defenders to inspect, test and improve.

NVIDIA is contributing open models, model weights, data and agent harness research to the Alliance. The new NVIDIA Labs Object-Oriented Agent framework is available on GitHub to make advanced AI safety capabilities more accessible for agent harnesses, enabling better integration between harnesses and models to make agent behavior easier to test, trace, audit and govern. Contributors across the Alliance are building an open defense stack for agents covering identity, isolation, safe model formats, multi-model scanning and secure coding workflows.

HPE contributes SPIFFE and SPIRE standards for zero-trust identity frameworks that cryptographically verify AI agents and services. Hugging Face has offered Safetensors to the PyTorch Foundation—a safe format for storing model weights that provides transparency and guarantees no remote code execution. IBM and Red Hat's Lightwell extends security across the open source supply chain with digitally signed patches. Microsoft's MDASH multi-model agentic scanning harness orchestrates specialized AI agents to discover, debate and prove exploitable bugs. SpaceXAI has open sourced the Grok Build terminal-based AI coding agent and plans to open source the weights of its model line to support developer and research communities.

As policymakers and regulators address AI safety, they must recognize open models, harnesses and security tooling as defensive assets, not liabilities, in AI policy. Blanket restrictions on open frontier AI systems would weaken defensive capacity and concentrate power and dependence in a few closed providers. Companies and governments should invest in shared open infrastructure for AI defense including datasets, evaluation frameworks, attack simulators and red-teaming tools, much as past generations invested in open source software itself.