Archive · 29 Sep 2026
Nvidia Launches Platform to Keep Autonomous AI Agents Safe
Nvidia has made a system to help keep computer programs called AI agents under control.
These agents can do tasks instead of only answering questions.
OpenShell puts an agent inside a protected software area with rules about what it can access.
It can control files, networks, tools and other computer services.
Sentry is a separate hardware-based watcher that checks what the agent is doing.
Nvidia says Sentry can isolate and stop an agent within milliseconds if it crosses its limits.
The company says more than 100 organisations are working with the platform.
The system is meant to add safety controls outside the AI model itself.
However, it cannot stop every mistake or poor decision an agent makes while using permissions it has been given.
Nvidia launched the Open Agent Safety Platform, combining OpenShell software with Sentry hardware monitoring.
OpenShell places AI agents in isolated runtimes and controls access to files, networks, tools, processes and credentials.
Sentry runs on BlueField-4 data processing units and can quarantine an agent in milliseconds, Nvidia says.
More than 100 organisations are reportedly working with the platform, including Anthropic, Salesforce, Microsoft and JPMorganChase.
The system is intended to contain agents but cannot prevent every mistake or bad decision within authorised boundaries.
- Who
- Nvidia announced the platform, with more than 100 organisations reportedly working with it, including Anthropic, Salesforce, SAP, Scale AI, Microsoft, Palantir and JPMorganChase.
- What
- The Open Agent Safety Platform combines OpenShell, a software boundary for AI agents, with Nvidia Sentry, hardware-based monitoring and enforcement.
- When
- Nvidia announced the platform on September 28, 2026, according to a post referenced in the article.
- Where
- The tools are designed for AI agents operating across software systems, including enterprise, coding and cybersecurity environments, and potentially robotics.
- Why
- Nvidia says independent controls are needed because increasingly autonomous agents can move beyond their intended tasks and should not be expected to police themselves.
Pacing Advocates
Layered-Control Advocates
Speed of AI development
Pacing Advocates
Anthropic CEO Dario Amodei has called for the pace of frontier AI development to be moderated so safety measures can keep up with advancing capabilities.
Layered-Control Advocates
Nvidia's approach is to add independent technical controls around AI agents while development continues.
Who should enforce safety
Pacing Advocates
The reported incidents and concerns suggest that relying on an AI system to follow instructions or govern its own behaviour may be insufficient.
Layered-Control Advocates
Nvidia argues that restrictions should sit outside the model, with OpenShell and Sentry enforcing boundaries independently.
Effectiveness of controls
Pacing Advocates
Technical boundaries cannot necessarily stop an agent from making a mistake or bad decision within the permissions it has been granted.
Layered-Control Advocates
Nvidia says multiple control layers can monitor activity, verify identities, enforce policies and stop agents that exceed their permitted boundaries.
- Platform
- NVIDIA Open Agent Safety Platform
- Software layer
- OpenShell provides an open-source runtime boundary that isolates agents, traces activity and enforces policies.
- Hardware layer
- Nvidia Sentry runs on BlueField-4 data processing units and monitors agents independently of their software environment.
- Enforcement speed
- Nvidia says Sentry can quarantine an agent in milliseconds if it crosses its permitted boundary.
- Policy coverage
- Sentry can inspect requests and responses, verify identities and apply access policies to data, tools, APIs and services.
- Processor support
- OpenShell is designed for Nvidia Vera CPUs and can be extended to Arm and Intel platforms.
- Key limitation
- The platform does not by itself prevent mistakes, deceptive behaviour or bad decisions made within an agent's authorised permissions.
