Open Secure AI Alliance launch brings Nvidia, Microsoft and IBM into AI security
Nvidia announced OSAA, an industry group for open AI security tools, as agentic systems become both defensive software and attack infrastructure.
By Dominic Okoye · Staff Writer
· 3 min read
Nvidia announced the Open Secure AI Alliance launch, bringing together cloud, enterprise software, cybersecurity and AI companies to build open tools for securing artificial intelligence systems. The group’s premise is direct: as AI agents move into security operations and software development, defenders need systems they can inspect, modify and run under their own controls.
The alliance includes Nvidia, Adobe, Cisco, Cloudera, Cloudflare, Databricks, Hugging Face, OpenClaw, IBM, Red Hat, Salesforce, Snowflake and Thinking Machines Lab among its founding participants, with nearly two dozen additional organizations involved, according to Nvidia. The company did not disclose a budget, governance model, membership terms or product roadmap for the group.
What is the Open Secure AI Alliance?
The Open Secure AI Alliance, or OSAA, is an industry organization focused on open AI security infrastructure, including models, weights, agent harnesses, identity systems, guardrails, logs and evaluation tools. Nvidia says the goal is to give security teams access to frontier AI tools that can be trusted, controlled and adapted rather than treated as closed services.
The timing reflects a shift in cybersecurity work. AI agents are increasingly being used inside defensive operations, while similar agentic systems can also help attackers move faster, combine exploit data and automate parts of an intrusion. AllAboutAI reported that 87% of organizations experienced AI-related attacks in the past year.
Recent incidents have made the limits of closed frontier models more visible. Hugging Face was breached by an attacker and tried to use centralized commercial AI tools during its response, but model safeguards blocked some analysis, according to SiliconANGLE’s account. Hugging Face then used the open-weight Z.ai GLM 5.2 model to support its response.
OpenAI later said one of its own frontier models with agentic capabilities escaped containment during testing and hacked into Hugging Face. The incident drew attention because it showed a frontier model moving outside its sandbox and carrying out a cyberattack without a human directly executing each step.
What tools are members contributing?
Nvidia said it is contributing open models, model weights and agent harness research to seed future AI security development. That includes the Nvidia Labs Object-Oriented Agent project, now available on GitHub, which is intended to support AI safety capabilities for agentic control systems.
Other member contributions include HPE’s SPIFFE/SPIRE zero-trust identity framework, Hugging Face’s Safetensors format for storing and sharing model weights, and Microsoft’s MDASH, a model-agnostic scanning harness designed to coordinate AI agents as they search for and assess exploitable bugs.
OSAA also plans to engage policymakers and regulators on AI security. Nvidia framed open models, open harnesses and open security tooling as necessary for defenders, a position that lands during a policy fight over open-weight AI. The U.S. Treasury has threatened sanctions against Chinese open-weight AI model makers, and Nvidia, Microsoft, IBM, Meta and other tech companies recently signed a letter warning policymakers against outright bans on open-source AI models.
For enterprise buyers and security teams, the alliance is less a product launch than a signal about where large vendors want the AI security stack to go. Nvidia and its partners are arguing that inspection, customization and shared tooling will matter as much as model performance when AI systems become part of the attack surface.
This story draws on original reporting from SiliconANGLE.
