NVIDIA has announced the Open Secure AI Alliance, a new industry group focused on open tools for AI safety and cybersecurity rather than another general-purpose model launch.

What changed

The company says the alliance will build and share open software that helps defenders inspect, adapt, and run AI security systems across their own infrastructure. NVIDIA's framing is policy-aware: open models and open harnesses are described as defensive assets, especially for organizations that cannot rely entirely on opaque hosted systems during incident response.

The announcement names founding members across cloud, security, enterprise software, open source, and AI research, including NVIDIA, Adobe, Cadence, Capital One, Cisco, CrowdStrike, Hugging Face, Microsoft, Red Hat, ServiceNow, and others. The practical focus is broader than model weights. NVIDIA argues that real AI safety depends on the agent stack around models: identity, permissions, harnesses, guardrails, logs, and evaluation.

Why it matters

The clearest developer artifact tied to the launch is NVIDIA Labs Object-Oriented Agents, or NOOA. NVIDIA describes it as an open-source research preview for building agents as Python classes, with methods as capabilities, fields as state, docstrings as prompts, and type annotations as contracts.

The NOOA repository says the framework includes optional packages for a CLI, long-term memory, and evaluation. It also warns that the software can execute LLM-generated code and should be run in isolated environments.

That caveat is important. The alliance does not make open agent systems safe by declaration. It does, however, put NVIDIA's weight behind a more inspectable security tooling layer, where defenders can test and modify the harnesses surrounding AI agents instead of treating them as sealed products.