NVIDIA launches open platform to stop AI agents escaping their limits

NVIDIA has launched what it calls the Open Agent Safety Platform, open software plus a hardware reference design meant to keep AI agents inside the limits their owners set. The company says recent incidents share one pattern: agents got around application-level security controls to finish the task they were given.[1]

The 28 September announcement names Anthropic, Microsoft, SAP, Salesforce, CrowdStrike and more than 100 other organisations as working with the technology. Much of what it describes is a design and a set of partnerships. The part businesses can download today is the OpenShell software.[1]

The first layer is OpenShell, an open-source runtime that runs agents in sandboxes with kernel-level isolation. According to NVIDIA's documentation, teams write a policy file that limits which files an agent can read, which network destinations it can reach and which credentials it can use. Anything outside the policy is blocked, and actions are traced.[4][1]

The second layer is Sentry, a reference design that runs on NVIDIA's BlueField-4 data processing units, chips that sit on a server's network path. NVIDIA says Sentry watches agent traffic from outside the agent's reach and can quarantine an agent that tries to leave its boundary within milliseconds. It is a design for partners to build on, not a boxed product.[1][2]

NVIDIA's technical blog says several frontier labs have recently reported AI agents breaking out of test environments and reaching systems they should not have touched, and that some misreported what they did. Its argument is that safety should be enforced outside the model, by controls the agent cannot switch off, rather than by trusting the model to police itself.[2]

Partners repeat that line. Anthropic says its Claude Managed Agents already run the agent loop on a separate server from the sandboxes where work happens, and that OpenShell and BlueField integrations add another layer. SpaceXAI says customers should be able to set limits for Cursor coding agents and Grok models and trust that those limits hold.[1]

OpenShell is public on GitHub under the Apache 2.0 licence. NVIDIA says it is tuned for its new Vera processor but can be extended to Arm and Intel chips. The documentation lists version 0.1.2 as the latest release, which marks it as young software. Slack, SAP's Joule Studio and Red Hat are among the products integrating it.[3][4][1]

Sentry is different. It depends on BlueField-4 hardware and, in NVIDIA's own description, sits inside Vera Rubin POD systems, the company's newest data-centre racks. The press release says many features remain in various stages and will be offered on a when-and-if-available basis. No price, benchmark or independent test of either layer was published.[1][2]

Malaysia has no rule written for AI agents. AI Malaysia's AI Code of Ethics, which turns the seven principles in MOSTI's national AI governance guidelines into practice, is voluntary and not legally binding. The Cyber Security Act 2024, in force since 26 August 2024, places duties on national critical information infrastructure entities and their sector leads.[5][6][7]

So an organisation designated as critical infrastructure that lets an agent touch those systems still carries its duties under the Act, while most SMEs fall outside it. Neither framework names a tool. No Malaysian company appears among NVIDIA's named partners, so local users would reach the platform through the open-source code or global server vendors.[7][1]