To rein in AI agents, Nvidia moves safety from the prompt to the silicon

According to NVIDIA, in recent incidents AI agents bypassed application-level safety controls in order to finish the tasks they had been given: hence the decision to go one layer deeper. On September 28, NVIDIA announced the Open Agent Safety Platform, an open software platform and reference system design for monitoring and containing autonomous systems from testing through deployment. The proposal shifts protection away from model guardrails and into the infrastructure: the architecture includes the open-source OpenShell software and a hardware reference design called Sentry.

According to the official GitHub repository, released under the Apache 2.0 license, OpenShell provides a kernel-level runtime boundary that controls file access, system calls and network connections, plus a credential management system that keeps real credentials hidden from the agent. It supports Linux, macOS (Apple Silicon) and Windows via WSL 2. The repository already has about 9,600 stars and 1,300 forks, a sign that what is new here is the overall package rather than the code itself. The company's press release describes it as optimized for NVIDIA Vera CPUs but extendable to Arm and Intel architectures. The Sentry component, built on NVIDIA DOCA software, is instead designed to run on NVIDIA BlueField-4 DPUs as an out-of-band supervisor that can isolate or quarantine agents when they violate policy. According to Infosecurity Magazine, quarantine can kick in within milliseconds, but the official documents do not say whether Sentry is open source, nor do they give a commercial availability date for BlueField-4 cards. So far there are no independent assessments of how effective the sandbox or the formal policy verification actually are.

The announcement ties into the Open Secure AI Alliance initiative backed by the Linux Foundation and names more than a hundred partner organizations, including Anthropic, Microsoft, Salesforce, SAP, JPMorganChase, Dell and Cisco. The company's statements do not specify, however, whether these names reflect deployments already in production or simply support in principle. “Security requires full-stack engineering. The Nvidia Open Agent Safety Platform brings together industry, researchers and public-sector organizations to share best practices,” said Jensen Huang, the company's founder and CEO.

Moving the agents' cage from the prompt layer down to system calls and hardware modules is a pragmatic choice. It answers a problem that has already surfaced: in recent incidents, application-level controls were not enough to stop agents determined to finish their task. What remains to be seen is whether the price of this protection is dependence on one very specific physical infrastructure. — Olya