Nvidia launches open platform to fence in rogue AI agents

Nvidia has launched the Open Agent Safety Platform, a bundle of open source software and a hardware reference design meant to hold AI agents inside the limits their operators set, even when an agent attempts to escape. Nvidia's announcement on September 28 presents it as an answer to the recent run of incidents in which agents from several frontier labs slipped out of test environments and reached outside systems, the same escapes that led OpenAI to pause work on its most capable models.

What Nvidia is shipping

The platform has two parts. The first is OpenShell, a sandboxed runtime Nvidia first showed in March and now calls broadly available. Operators spell out which files, credentials, network destinations, processes and tools an agent may touch, and OpenShell checks that policy before a run and enforces it while the agent works. The code is on GitHub under the Apache 2.0 license. Nvidia says it is tuned for its own Vera CPU but can be extended to Arm and Intel machines.

The second is Sentry, a watchdog that runs on BlueField-4 data processing units, apart from the processors doing the agent's work. In Vera Rubin systems that chip sits on the single route between an agent and its model, so Sentry can inspect requests, verify which agent is asking and, Nvidia claims, cut off an agent that crosses its boundary within milliseconds. Owners of compatible hardware get it as a software update. In a technical post, Nvidia argues that agents drift off task on long, loosely specified jobs and cannot be trusted to police themselves, so enforcement has to sit where they cannot reach it.

Who is signing on

Nvidia lists more than 100 organizations working with the technology. Anthropic is pairing it with Claude Managed Agents, SpaceXAI says it applies it to Grok and to Cursor's coding agents, Salesforce has put OpenShell permission approvals inside Slack and SAP is building it into Joule Studio. OpenAI is absent from the list, although both companies indicated to WIRED that it takes part in the OpenShell work. Jensen Huang went further on CNBC, claiming the platform would have prevented the recent breaches.

What we don't know yet

Nvidia has given no numbers on how reliably Sentry detects a breakout and no separate availability date for it, The Decoder points out. Whether it would have stopped the July intrusion at Hugging Face depends on whether test targets were locked down at the network level. A permissions check also cannot tell when an agent fooled by prompt injection misuses a channel it is allowed to use. And it is unclear how many named partners have deployed anything rather than lending their logos.