Sentry adds another layer of monitoring. The software runs on Nvidia’s BlueField programmable data processing units and continuously watches agents as they operate. If an agent attempts to move beyond its assigned boundaries, Nvidia says Sentry can intervene and quarantine it.
“It can quarantine a suspicious agent in milliseconds,” Justin Boitano, Nvidia’s vice president of enterprise AI, said.
The release comes after several incidents involving autonomous agents behaving in unexpected ways. OpenAI recently disclosed that some of its agents interacted with U.S. government websites in ways the company did not anticipate, while another incident involved OpenAI agents accessing AI company Hugging Face.
Nvidia executives argue that traditional software sandboxes are not sufficient for organizations deploying large numbers of autonomous agents. Boitano said customers increasingly want to operate fleets of agents, creating a need for security policies that can be enforced across all of them.
Nvidia says OpenShell allows developers to “formally verify an agent has enough authority to do its job and no more,” according to Boitano.
The company is also working with Arm and Intel to expand Sentry beyond Nvidia hardware and bring it to the x86 architecture.
“Once it runs on those instruction-set architectures, it can run on any architecture,” Boitano said.
Nvidia says more than 100 organizations are using the platform at launch. Its broader group of AI safety collaborators includes Anthropic, Cisco, CoreWeave, CrowdStrike, Dell Technologies, Hugging Face, JPMorganChase, Microsoft, Mistral and Palantir.
SpaceXAI is using the platform with its Cursor agents and Grok models, while Nvidia and Anthropic are working together on security for Claude Managed Agents. Salesforce, Scale AI and SAP are also integrating OpenShell to some extent.
OpenAI does not appear on Nvidia’s publicly disclosed partner list, although both companies indicated that OpenAI is involved with the OpenShell effort. Neither provided an explanation for the omission.
The Open Agent Safety Platform also connects with a broader Nvidia-led AI safety initiative launched in July. That coalition now includes more than 120 companies and operates the Shared AI Findings Exchange, or SAFE, which is intended to distribute information about AI safety issues across participating organizations.
Nvidia says that program is designed to operate independently rather than being controlled by one company or industry group.
The company is positioning hardware-enforced restrictions as a way to make increasingly autonomous agents safer to deploy. Instead of relying solely on an agent to follow instructions, OpenShell and Sentry are designed to enforce those restrictions separately and intervene when the software attempts to cross them.
This analysis is based on reporting from CBS News.
Image courtesy of Nvidia.
This article was generated with AI assistance and reviewed for accuracy and quality.
