NVIDIA has launched the Open Secure AI Alliance, a group of more than 120 companies building and sharing open technologies for AI safety and cybersecurity. Arcade.dev has joined as a founding partner.
We joined because the core premise matches what we have argued since we started building. Real AI safety and security depend on the full agent stack — identity, permissions, harnesses, guardrails, logs, and evaluation — not just on whether a model’s weights are open or closed.
Safety holds or breaks below the model
The industry largely spent the last few years on the reasoning layer, on how agents think. That work matters, but an agent that only thinks accomplishes nothing on its own. The moment the agent acts, closing a ticket or updating Salesforce, it has to authenticate, get authorized, and stay inside its organization’s governance rules.
Enterprises learned a version of this before agents existed. They adopted IAM services for identity and kept them independent from the systems they unlocked. The same logic applies to agents. Few companies will let one provider own both the model and the pathway its agents use to act. When the layers stay separate and the action layer stays open, security teams get a stable place to enforce policy, one that holds even as the models underneath it change.
NVIDIA is opening its models and separating the agent harness into its own layer. We read that as a welcome move for an ecosystem that works better when no single layer is a black box.
We run this layer in production
Arcade is the actions runtime for enterprise AI agents. We enforce security policies on every action, execute reliably across business systems, and give teams one place to govern agents in production. For the past two years, we’ve run that layer inside enterprises, at scale, and under real regulatory requirements.
That experience is our primary contribution to the alliance. Authorization evaluated per action, at the intersection of user identity and agent scope, with no standing service-account access. Token lifecycles handled by the runtime and kept away from the model. Every action logged to your security information and event management (SIEM) system through OpenTelemetry, so a prompt injection becomes a logging event rather than a breach.
These practices run in production today, and we want to help harden them into shared, open standards.
We stay agnostic as the standards evolve
New models arrive constantly, some open and some closed, alongside new frameworks, clients, and standards. Arcade works across them all today, and will keep working across whatever comes next. This agnosticism is what serves our enterprise customers best, and where a strong enterprise standard can take hold, we help push it forward.
That is also why joining made sense. An alliance built on open, shared tooling gives the enterprises we serve more portability and less lock-in, exactly what a security team wants before it trusts agents in production.
We give developers open frameworks to build, orchestrate, and observe agents on any model. Then Arcade covers what happens when an agent acts, enforcing authorization and security on every action. Keeping those frameworks and action layers open and independent keeps the whole ecosystem flexible, which is what agentic workflows need to scale.
The voice we bring to the alliance
We will keep representing the voice of the enterprise customer. The teams deploying agents to check emails, write code, and process sensitive contracts carry unusually hard security and governance requirements, and they set the bar for anything we help standardize, from the Model Context Protocol (MCP) to token refresh and authorization.
No single company finishes security on its own. It holds up when the ecosystem treats it as shared work. We are glad to contribute what we have learned, and to keep learning from everyone we’re building alongside.
To learn more, read NVIDIA’s announcement of the Open Secure AI Alliance.
