NVIDIA brings industry partners together in an Open Secure AI Alliance

NVIDIA and a broad group of technology companies have formed the Open Secure AI Alliance, putting open-source security tools at the centre of a new industry effort. The founding group includes infrastructure providers, cybersecurity companies, enterprise software vendors and open-source organisations. Its stated aim is to develop and share technologies that help defenders secure software and AI agents as those systems become part of everyday business and public infrastructure.

The announcement comes as the attack surface around AI expands beyond the model itself. A system can be exposed through training data, a model-serving endpoint, a plug-in, a dependency or an application that grants the model access to company information. Agents add another layer because they can read documents, call APIs, use credentials and make several decisions before a person reviews the result. Traditional application-security controls remain necessary, but they do not by themselves explain what an agent is allowed to do or whether its actions can be traced.

The alliance says it will build on open-source security work, including projects associated with the Linux Foundation and the Open Source Security Foundation. NVIDIA is contributing models, weights, data and research on agent harnesses. One early project from NVIDIA Labs is an object-oriented agent framework intended to make agent behaviour easier to test, trace, audit and govern. Other participating organisations are contributing identity standards, safer model-weight formats, signed patches and multi-model scanning tools.

Openness matters because many organisations need to inspect a security control, adapt it to local systems or run it without sending sensitive data to a third-party service. Shared tools can also give smaller engineering teams access to tests and monitoring components that would be expensive to build alone. The alliance is making an argument that defensive capability should not be concentrated in a few opaque products, particularly when a vulnerability can affect a large number of downstream users.

That approach also creates difficult coordination questions. Open projects need maintainers, funding, release processes and a clear way to disclose and remediate vulnerabilities. Security guidance must cover the full agent stack, including identity, permissions, isolation, tools, logs and evaluation, rather than focusing only on model weights. A control that protects a model while leaving an agent's credentials or an unpatched dependency exposed would offer limited protection in a real incident.

For companies deploying AI, the initiative could provide a common vocabulary for assessing systems that combine models with software and external actions. Its influence will depend on whether the alliance releases code and test methods that work outside a single vendor's platform and whether those projects remain maintained after the launch. The announcement is therefore better understood as the start of shared security infrastructure than as a finished product.

The first meaningful test will come when defenders can run the alliance's tools during an actual incident and compare their results across different models and cloud environments. As AI systems gain more ability to act, the quality of the surrounding security ecosystem will matter as much as model capability. Open collaboration could make that ecosystem more resilient, but only if the participating companies turn the promise of transparency into software that is reliable, auditable and regularly updated.