As artificial intelligence agents become capable of performing increasingly complex tasks on their own, controlling what they can access and execute has become a major technology challenge. NVIDIA has now introduced a new security platform designed to keep these autonomous systems within clearly defined boundaries.
NVIDIA Introduces Open Agent Safety Platform
NVIDIA announced its Open Agent Safety Platform on September 28, bringing together open-source software and hardware-based security tools to help organisations monitor and control AI agents. The platform is designed to provide protection from testing through deployment.
The company says the system can help prevent AI agents from moving beyond the permissions given to them.
OpenShell Sets Boundaries for AI Agents
One of the platform’s main components is OpenShell, an open-source secure runtime. It tracks an agent’s actions and enforces policies while the agent performs tasks.
The idea is to ensure that an AI system receives enough authority to complete its assigned job without gaining unnecessary access to files, networks, tools or other resources.
Because OpenShell is open source, NVIDIA says it can also be extended to work with computing platforms from companies such as Arm and Intel.
Sentry Adds a Hardware Security Layer
The second major component is Sentry, a security watchdog designed to independently monitor AI-agent behaviour.
According to NVIDIA, Sentry runs on its BlueField-4 data processing units and can detect when an agent attempts to move outside its permitted boundaries. The system can then quarantine the suspicious agent within milliseconds.
This creates two layers of protection: OpenShell controls what the agent is allowed to do, while Sentry provides independent monitoring and containment.
Why AI Agent Security Matters
AI agents are increasingly being used to perform tasks that involve computer systems, data and online services. Recent incidents involving AI systems bypassing intended boundaries have increased attention on how these tools should be secured.
NVIDIA executives said the new platform could potentially have prevented a recent incident involving AI agents that autonomously accessed Hugging Face systems. That is NVIDIA’s assessment rather than an independently established finding.
More Than 100 Organisations Involved
NVIDIA said more than 100 organisations were using the platform at launch, including Microsoft, Perplexity, Accenture and JPMorgan Chase. The company has also highlighted collaboration across the wider AI ecosystem, including technology and cybersecurity companies.
A New Layer of AI Protection
The launch reflects a broader shift in AI security: instead of relying only on the model itself to behave safely, companies are increasingly looking at protections around the model and the computing environment in which an agent operates.
NVIDIA’s platform is aimed at providing those additional safeguards as autonomous AI becomes more capable and more widely deployed. Experts have also noted that challenges remain, including determining exactly what permissions an AI agent should receive while still allowing it to perform useful work.