A long partner list, and one piece of working code
A large group of AI, cloud and cybersecurity companies has formed the Open Secure AI Alliance.
NVIDIA announced the initiative on 27 July with the Linux Foundation and dozens of partners. The list spans Microsoft, IBM, Red Hat, Hugging Face, Cloudflare, Cisco, CrowdStrike, Databricks, Mistral, Mozilla, GitHub and many others that usually compete with one another.
The stated goal is to build open tools for securing AI software and agents. That can include identity, permissions, isolation, model formats, vulnerability scanning, secure coding workflows, logs and evaluation.
Alliances often arrive with a broad promise and little that can be used on day one. This launch has at least one concrete artifact: NVIDIA has released its Object-Oriented Agent framework, or NOOA, under the Apache 2.0 licence.
That does not make a security stack. It gives the announcement something testable, which is a better place to start than a list of logos alone.
What NOOA actually does
NOOA is a model-agnostic Python framework for building AI agents. It brings an agent's state, capabilities, prompts and typed inputs together inside a Python class.
Methods with ordinary code remain deterministic. Methods left for the language model become generated actions at runtime. Type annotations define the shape of inputs and outputs, while docstrings help tell the model what to do.
The approach is meant to make an agent look more like software that Python developers already know how to read, test, refactor and version. The repository also includes tracing, a development viewer, examples and optional evaluation and memory packages.
NVIDIA and the Linux Foundation describe those features as a way to make agent behaviour easier to test, trace, audit and govern. That is plausible, but the project is research software at an early stage. A framework can expose behaviour without proving that the behaviour is safe.
The useful question is not whether NOOA makes agents secure by itself. It is whether shared, inspectable plumbing helps teams find and fix failures before they become production incidents.
The repository carries the important warning
NOOA can let a model write and execute Python. Its own documentation is direct about the risk.
Generated code may send private data elsewhere, delete files or change its environment. NOOA performs syntax-tree checks and blocks some modules, but the maintainers say those measures are defence in depth, not a containment boundary.
The project tells users to put code-running agents inside operating-system-level isolation such as a container or virtual machine. Static checks inside the Python process cannot reliably stop code that is trying to escape.
That warning is not a reason to dismiss the project. It is a useful example of the problem the alliance says it wants to solve. Agent security is not just a safer model response. It includes credentials, file access, networks, logging, approvals and the boundary around the process itself.
Open code makes that design inspectable. It does not remove the need for careful deployment or an independent security review.
Shared infrastructure could help, if governance follows
The alliance is trying to borrow an old open-source idea: competitors can share the boring security infrastructure that none of them benefits from rebuilding badly in private.
NVIDIA points to work that already exists across its partners, including identity standards, safer model-weight formats, signed software patches and multi-model vulnerability scanners. The Linux Foundation says it can provide a neutral place for competing organisations to collaborate.
There is real value in interoperable logs, common test formats and coordinated vulnerability disclosure. A security tool that only understands one provider's agent stack is less useful in a company that uses five of them.
But the launch does not yet provide a detailed governance model, technical roadmap, funding structure or timetable. It is also not clear which projects will become shared standards, which will remain vendor-led, and how independent researchers will shape decisions.
A wide membership list signals interest. Delivery will be measured in maintained code, public issue handling, reproducible evaluations and standards that work outside the founding companies' own products.
What is confirmed, claimed and still open
Confirmed: the Open Secure AI Alliance was announced on 27 July with dozens of inaugural partners. NVIDIA released NOOA publicly on GitHub under Apache 2.0, and the Linux Foundation says it is joining the effort.
The partners' claim: open models and open harnesses give defenders more visibility, local control and freedom to adapt tools during an incident. They argue that shared infrastructure can reduce dependence on a few closed providers.
Still open: who makes final alliance decisions, how vulnerabilities will be disclosed, what common specifications will emerge, how progress will be measured and whether competitors will keep contributing after the launch moment.
Also confirmed: NOOA is not a sandbox. Its own maintainers say agents that execute generated code need a real isolation boundary around them.
That may be the most credible part of the launch. The alliance is not starting from the idea that one guardrail solves agent security. It is starting with the messier truth that the whole stack has to be visible and constrained.
Sources
- NVIDIA - Open Secure AI Alliance announcementPrimary announcement published 27 July 2026. Source for the partner list, stated mission, planned security layers and NVIDIA's initial contributions.
- Linux Foundation - joining the Open Secure AI AlliancePrimary partner statement used to verify the Foundation's role, the case made for shared open infrastructure and the limits it places on openness alone.
- NVIDIA NeMo - NOOA source repositoryPrimary code and documentation. Used to verify the framework design, Apache 2.0 licence, installation status, tracing features and explicit sandboxing warning.
