NVIDIA has joined about 40 technology organisations to establish the Open Secure AI Alliance, a new industry effort focused on open tools for protecting software and AI agents. The inaugural group spans cloud infrastructure, cybersecurity, enterprise software, model development and open-source foundations, giving the initiative a broad base from which to address risks across the agent stack.
The alliance is intended to develop and share technologies, techniques and tools that defenders can inspect, adapt and deploy. NVIDIA says the work will build on the Linux Foundation's Akrites initiative and the Open Source Security Foundation community, rather than starting as an isolated vendor program. Its central argument is that defenders need transparent systems they can operate within their own environments, alongside closed frontier services.
A broad membership with a security focus
Inaugural partners named by NVIDIA include Adobe, Cadence, Capital One, Cisco, Cloudera, Cloudflare, Cognition, CrowdStrike, Databricks, Dell Technologies, DoorDash, Elastic, HPE, Hugging Face, IBM, LangChain, Microsoft, Palantir, Palo Alto Networks, Red Hat, Salesforce, SAP, ServiceNow, Siemens, Snowflake and Synopsys. The Linux Foundation is also part of the group.
That range matters because an AI agent is more than a model. Production systems combine model weights with an execution harness, identity and access controls, tools, logs, guardrails and evaluation. A weakness in any one layer can change what the overall agent is able to do. The alliance says its contributors are therefore working across identity, isolation, safe model formats, multi-model scanning and secure coding workflows.
The announcement points to existing projects as examples of the components that could form an open defence stack. These include HPE's work on SPIFFE and SPIRE for workload identity, Hugging Face's Safetensors model-weight format, IBM and Red Hat's Lightwell supply-chain security work, and Microsoft's MDASH multi-model scanning harness. These projects remain distinct contributions; NVIDIA did not announce a single integrated product or an availability date for a complete stack.
NVIDIA contributes a new agent-harness project
NVIDIA says it will contribute open models, model weights, data and agent-harness research. Its most concrete new contribution is the NVIDIA Labs Object-Oriented Agent project, known as NOOA, which is now available as open-source research on GitHub.
NOOA is designed to improve how an agent harness integrates with models so behaviour can be tested, traced, audited and governed. The harness is the software layer that manages a model's interaction with tools and its surrounding system. Making that layer observable is important for security reviews because model evaluations alone do not show every permission, tool call or state change that may occur when an agent acts.
The practical value will depend on adoption, documentation and how well the framework works with different models and deployment environments. NVIDIA describes NOOA as a research framework, so teams should distinguish its current status from a supported enterprise security product. The announcement does not specify a commercial support model or compatibility matrix.
Why openness is central to the proposal
NVIDIA's case for the alliance is that open models and harnesses allow defenders to examine and customise their systems, keep sensitive incident data under their control, and avoid dependence on a single provider. It also acknowledges that capable open models can be misused or modified to weaken safeguards. The company argues that those risks should be addressed through strong controls, evaluation and rapid remediation rather than by denying defenders access to inspectable tools.
This is a policy position as well as a technical one. The announcement urges governments and companies to treat open models, harnesses and security tooling as defensive assets and to invest in shared datasets, evaluations, attack simulators and red-team infrastructure. It cautions against blanket restrictions on open frontier systems.
For enterprises, the immediate implication is not a new security product to buy. It is the emergence of a large collaboration that could shape common building blocks and operating practices for agent security. Open components can improve portability and scrutiny, but they do not remove the need for access control, monitoring, incident response and independent validation inside each deployment.
Questions the alliance still needs to answer
The launch statement establishes a mission and membership but leaves important implementation details open. It does not set out a formal governance structure, a roadmap, contribution requirements or a timetable for the first alliance-wide deliverables. It also does not explain whether all outputs will use consistent licences or how projects will be assessed before organisations rely on them.
Those details will determine whether the initiative becomes a durable open-security program or a loose collection of related projects. Prospective users should evaluate each released component on its own merits while watching for published governance, threat models, test results and maintenance commitments from the alliance.