NVIDIA, Microsoft, SpaceX, and other major firms have launched the Open Secure AI Alliance to develop open-source tools for defending software, AI agents, and infrastructure from cyberattacks, highlighting new security challenges as AI systems proliferate
Several leading technology and cybersecurity companies, including NVIDIA, Microsoft, SpaceX, and Adobe, have established the Open Secure AI Alliance, a new industry group focused on developing open-source tools to defend software, AI agents, and critical infrastructure against cyber threats. The alliance brings together organizations from cloud computing, enterprise software, cybersecurity, and open-source foundations, with founding members such as Capital One, Cisco, Cloudflare, CrowdStrike, Dell Technologies, Hugging Face, IBM, LangChain, the Linux Foundation, NAVER, NetApp, Palantir, Palo Alto Networks, Red Hat, Salesforce, SAP, ServiceNow, Siemens, SK Telecom, Snowflake, Synopsys, and others. Notably, Google, OpenAI, and Anthropic are not among the initial participants.
The alliance aims to address the growing security risks associated with the rapid deployment of AI systems by creating and sharing open technologies, techniques, and tools. According to the companies, open models and agent frameworks allow defenders to inspect, adapt, and deploy security solutions on their own infrastructure, reducing dependence on proprietary systems and single vendors. This approach is intended to give organizations greater control over sensitive data and security operations, especially as AI agents become more integrated into software and infrastructure.
Open Models in Cyber Defense
The alliance's formation follows recent incidents that exposed the limitations of closed AI systems in cybersecurity. In one case, Hugging Face reported using an open-weight GLM 5.2 model on its own infrastructure to analyze over 17,000 actions and contain an intrusion, after closed AI tools failed to distinguish between attacker and defender activity. The group argues that such incidents demonstrate the practical need for defenders to access advanced open systems, particularly when rapid response is critical.
While open models can be inspected and modified, the alliance acknowledges that they also carry risks of misuse, including potential use in cyberattacks or attempts to bypass safeguards. However, the group maintains that these risks are not unique to open systems and exist in closed models as well. The alliance advocates for pairing openness with strong safeguards, clear rules against malicious use, rigorous evaluation, and rapid remediation processes.
Technical Contributions and Security Stack
Members are contributing a range of technologies to the alliance. NVIDIA is providing open models, model weights, datasets, and research on agent harnesses, including the open-source NVIDIA Labs Object-Oriented Agent (NOOA) project. NOOA is designed to help agent harnesses interact with models in ways that make agent behavior easier to test, trace, audit, and govern. HPE is contributing to SPIFFE/SPIRE, which establishes zero-trust identity standards for verifying agents and services. Hugging Face has offered its Safetensors format for secure model weight storage to the PyTorch Foundation. IBM and Red Hat are working on digitally signed patches for the open-source software supply chain. Microsoft is contributing its MDASH multi-model agentic scanning harness, which uses specialized agents to discover and validate exploitable vulnerabilities. SpaceXAI has open-sourced the Grok Build terminal-based coding agent and plans to release the weights for the Grok model line.
The alliance emphasizes the need for defensive tools that can operate across multi-vendor ecosystems, aiming to reduce reliance on individual providers and avoid single points of failure. The group is also calling on policymakers to recognize open models, agent harnesses, and security tooling as essential defensive assets, warning that blanket restrictions on open systems could weaken cyber defenses and increase dependence on a small number of closed providers.
Deployment, Limitations, and Policy Context
At present, the Open Secure AI Alliance is focused on developing and sharing open-source tools and frameworks rather than deploying a single unified system. The initiative is in its early stages, with contributions from member organizations spanning model weights, agent harnesses, identity standards, secure storage formats, and vulnerability scanning tools. There is no evidence yet of independent evaluation or systematic benchmarking of the alliance's tools in operational environments. The effectiveness of open models in real-world cyber defense remains subject to ongoing testing and institutional adoption.
The alliance's approach raises unresolved questions about the balance between openness and security. While open models can improve transparency and adaptability for defenders, they may also lower barriers for attackers seeking to exploit vulnerabilities or remove safeguards. The group's position is that robust governance, evaluation, and rapid response mechanisms are necessary to mitigate these risks. The alliance is urging companies and governments to invest in shared infrastructure, including datasets, evaluation frameworks, attack simulators, and red-teaming tools, to strengthen collective security as AI agents become more widely deployed.
In a recent incident cited by the alliance, Hugging Face used an open-weight model to analyze and contain a security breach after closed AI tools proved inadequate. The company reported analyzing more than 17,000 actions during the investigation. This example is presented as evidence that open models can provide defenders with critical capabilities not always available in proprietary systems, particularly when rapid inspection and adaptation are required.
As the alliance develops, its impact will depend on the adoption and effectiveness of its open-source tools, the willingness of organizations to integrate them into existing security operations, and the regulatory environment governing open and closed AI systems. The absence of major AI developers such as Google, OpenAI, and Anthropic from the founding group highlights ongoing divisions within the industry over the role of openness in AI security.
Understanding the distinction between open and closed AI models is central to the alliance's strategy. Open models are those whose architecture, weights, and often training data are made publicly available, allowing independent inspection, modification, and deployment. This contrasts with closed models, where only the provider can access or modify the underlying system. In cybersecurity, open models can enable defenders to adapt tools to specific threats and infrastructure, but they also require careful governance to prevent misuse. The debate over open versus closed models reflects broader tensions in AI development between transparency, security, and commercial control.
