Bitdeer AI Collaborates with NVIDIA on the NVIDIA Open Agent Safety Platform to Build the Security Foundation for the Agentic Era

As AI agents move beyond simple chat interfaces to deeply integrate into enterprise business systems and gain autonomous decision-making and execution capabilities, agent safety has emerged as a core mission for the AI industry. As enterprises pursue maximum efficiency, they face a critical question: How can they ensure that agents, while equipped with autonomous problem-solving capabilities, remain strictly within their approved security sandboxes and permission boundaries?

At Bitdeer AI, we believe that empowering agents with autonomy must be anchored in rock-solid security. Security and compliance aren't the brakes that slow innovation down, they're what makes speed possible. That’s why Bitdeer AI is working with NVIDIA on the launch of  NVIDIA Open Agent Safety Platform. We will integrate NVIDIA's AI safety technology stack to build an enterprise-grade security framework providing both software- and hardware-isolated protection for AI agents.

I. The Breakthrough: Shifting from "Model Self-Constraint" to "Infrastructure-Level Enforcement"

Traditional prompt engineering and model-level application guardrails have proven insufficient in complex agentic scenarios. When an agent is empowered to solve problems creatively, it also gains the potential to find unexpected paths or exploit vulnerabilities to bypass instruction limits.

We believe the safer paradigm is one in which security policies are enforced by the underlying infrastructure, rather than relying on the agent's own judgment. Even if an agent misunderstands a task, hallucinates, or falls victim to malicious prompt injection attacks, as long as the infrastructure layer enforces hard physical and network boundaries, the agent cannot step outside its authorized perimeter.

II. Full-Stack Architecture: Software and Hardware-Integrated NVIDIA Open Agent Safety Architecture

The NVIDIA Open Agent Safety Platform brings together software controls and hardware-backed protections to make autonomous AI agents safer to run at enterprise scale. The platform delivers that in two key pillars:

NVIDIA OpenShell is the secure runtime that governs how the agent executes, what it can see and do, and where inference goes. It creates the boundary that separates agents’ work from how they interact with their environment and the rest of the world.

Features: Individual agent sandboxes, a policy enforcement engine and a gateway govern access to files, networks, credentials, tools and model endpoints. Policy Prover adds checks for policy violations or sandbox escapes; allow/deny logs support audit.

NVIDIA Sentry adds an out-of-band watchdog that runs on NVIDIA BlueField-4 DPUs to continuously monitor agent behavior. Sentry can quarantine agents that attempt to move outside their boundaries in milliseconds. 

OpenShell is Apache 2.0-licensed, can run on Vera CPUs, and integrates seamlessly with existing identity providers, credential stores, observability, and security tools. This builds on NVIDIA’s open model commitment and the Open Secure AI Alliance. As an NVIDIA Cloud Partner, Bitdeer AI brings these verifiable, inspectable protection layers into our platform services through our own confirmed capabilities, making trust boundaries completely transparent.

III. Zero-Trust in Practice: Bitdeer AI's Sandboxed Production Workflows

On Bitdeer AI Cloud, we are practicing Zero-Trust architecture by fully implementing OpenShell protections. Bitdeer AI is listed on build.nvidia.com as one of the first inference providers for agents running in OpenShell sandboxes, and we have published a guide to deploying sandboxed NemoClaw environments on Bitdeer AI Cloud. 

Consider a typical enterprise scenario where an agent works with private, sensitive company information. An OpenShell policy on Bitdeer AI Cloud delivers complete end-to-end security:

●      Least-Privilege Access: The agent has read-only access to an approved internal document system while all public internet access is blocked.

●      Privacy-Locked Inference: Its model calls are pinned by the privacy router to Bitdeer inference endpoints, ensuring sensitive context does not flow to unapproved third-party services.

●      Tamper-Proof Permission Boundary: Its filesystem and process permissions are locked when the sandbox is created, preventing it from granting itself new access or installing malicious dependencies mid-task.

●      Full Auditability: Every allow and deny decision is logged by the infrastructure to support customer audits.

To lower the barrier to adoption for enterprises, Bitdeer AI provides optimized OS images and policy configuration recipes, allowing customers to quickly deliver compliant, secure sandboxes simply by configuring policy parameters.

At the same time, on our hardware roadmap, we are actively evaluating and planning the deployment of NVIDIA BlueField-4 and NVIDIA Vera CPU architectures. This will extend network enforcement beyond the host, provide an extra layer of chip-level threat detection and isolation, and supply high-performance compute for CPU-intensive agent orchestration.

Notably, Bitdeer AI is not only an enabler of agent safety but also an active practitioner of agentic AI in real workflows. Within our own enterprise operations, we have established a full-lifecycle, multi-dimensional security governance framework, spanning tool onboarding, endpoint protection, network monitoring, relay auditing, and correlated response, ensuring that our internal agentic workflows operate reliably within approved boundaries.

IV. Open Collaboration: Building Verifiable Trust for the Future of Scalable Agents

Security must never be a black box. Uninspectable security is merely a promise, whereas security built in open environments provides verifiable assurance. Open development enables the entire industry to inspect control measures, actively test failure scenarios, and share valuable lessons learned, thereby elevating the safety and resilience of the agent ecosystem as a whole.

Bitdeer AI strongly shares NVIDIA’s strategic vision of advancing AI safety in open environments in the agentic era. Moving forward, we will continue to support and collaborate closely with NVIDIA and ecosystem partners to build secure, scalable and trustworthy AI cloud infrastructure, powering agentic AI workflows to thrive safely in real production environments.