Open-source software is recognized as a foundational element supporting the global economy, providing accessibility and transparency across sectors such as cloud computing, finance, manufacturing, telecommunications, government services, and internet infrastructure. Cybersecurity has emerged as one of the primary beneficiaries of this open technology.
The Mandate for Open AI Security
The newly established Open Secure AI Alliance (OSAA) aims to proactively address and disclose vulnerabilities by utilizing open technologies. Building upon the existing work done by the Linux Foundation’s Akrites initiative and the OpenSSF community, the alliance is responding to a critical choice in artificial intelligence security for the United States and its allies: whether defensive protections will be contained within a small number of proprietary (opaque) systems or built upon openly available models, tools, and frameworks that any defender can thoroughly study, adapt, and implement.
While acknowledging that both closed and open technological models have their uses, the alliance emphasizes that for cybersecurity applications, open models and harnesses are essential. They help democratize defensive capabilities, enhance transparency for security professionals, allow defense while simultaneously protecting sensitive data, and complement advanced proprietary systems with highly customizable, localized controls. Open source facilitates a massively distributed, community-driven, and self-controlled approach to defense, thereby eliminating single points of failure.
Addressing AI Security Risks
The necessity for open systems was highlighted by recent security events, such as the incident at Hugging Face. During that event, when closed AI tools were unable to differentiate between legitimate defenders and attackers, they obstructed essential forensic analysis. To contain the intrusion and analyze over 17,000 actions, Hugging Face utilized the open-weight GLM 5.2 model running on its own infrastructure.
This incident demonstrated a crucial point: when security teams lack the ability to inspect, modify, and operate advanced AI systems internally, their capacity for rapid response is severely limited precisely when speed is paramount. Therefore, critical sectors require open frontier defensive tools and techniques to build secure systems across complex multi-vendor environments and avoid reliance on any single provider.
Alliance Participation and Scope
The core mission of the Open Secure AI Alliance is ensuring that security practitioners globally have access to trustworthy, controllable, and cutting-edge open tooling. The alliance has secured participation from a wide array of industry leaders across cloud computing, enterprise software, open source foundations, and AI research. 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.
Technical Advancements in Defense Stacks
The alliance focuses on the entire “agent stack”—encompassing identity, permissions, harnesses, guardrails, logs, and evaluation—recognizing that AI safety depends on more than just whether model weights are open or closed. The following contributions showcase how partners are building an open defense infrastructure for agents:
- NVIDIA is contributing open models, associated data, and new agent harness research to accelerate the development of cybersecurity tools. They have released the new open-source NVIDIA Labs Object-Oriented Agent (NOOA) project on GitHub, which helps integrate agent harnesses with models, making agent behavior easier to test, audit, and govern.
- HPE contributes to SPIFFE/SPIRE, a framework that establishes zero-trust identity standards. This system allows for cryptographic verification of AI agents and services, ensuring only authorized workloads can access enterprise resources.
- Hugging Face has provided Safetensors—a secure format for storing AI model weights—to the PyTorch Foundation. This format guarantees transparency and prevents remote code execution risks.
- IBM and Red Hat developed Lightwell to extend security across the open source supply chain by implementing digitally signed patches.
- Microsoft introduced MDASH, a multi-model agentic scanning harness designed to orchestrate specialized AI agents that can discover, debate, and prove exploitable software bugs.
- SpaceXAI has released the Grok Build terminal-based AI coding agent to promote transparency. The company also plans to open source the weights for its entire Grok line of models to support research communities.
Policy Recommendations
The alliance strongly urges policymakers and regulators to recognize open models, harnesses, and security tooling as defensive assets, rather than viewing them as liabilities in AI policy. The group argues that imposing blanket restrictions on open frontier AI systems would weaken defense capabilities and risk concentrating power within a few proprietary providers.
Instead, governments and corporations should invest in shared, open infrastructure for AI defense—including attack simulators, evaluation frameworks, datasets, and red-teaming tools—mirroring historical investments in general open source software. By adopting this collaborative approach, the industry aims to ensure that the safety and security of advanced AI are built openly, promoting resilience, competition, and technological leadership for everyone.
