Nvidia makes sandbox OpenShell available and introduces agent monitoring tool Sentry
Nvidia has released the security sandbox OpenShell to the public and introduced Sentry software for monitoring AI agents on Bluefield DPUs. Dozens of partners have signed on to the project, including Anthropic, Microsoft and Mistral, but OpenAI is missing.
Nvidia has made the security sandbox OpenShell generally available, a framework first announced in March at the GTC conference. OpenShell isolates the activity of AI agents at the operating system kernel level, the fundamental layer that has access to practically all parts of a computer system. At the same time, Nvidia introduced new software called Sentry, an isolated security domain intended primarily for Bluefield-series chips (programmable DPUs), which is meant to continuously monitor long-running agents and, according to the company, can "quarantine" agents that attempt to step outside established boundaries. Nvidia is also working with Arm and Intel on a version of Sentry for the x86 architecture. Both tools now fall under an umbrella framework called the Open Agent Safety Platform.
According to Nvidia's materials, the company has security collaborations with dozens of other technology companies, including Anthropic, Cisco, CoreWeave, CrowdStrike, Dell Technologies, Hugging Face, JPMorganChase, Mistral, Microsoft and Palantir. Nvidia states that SpaceXAI uses the platform for Cursor agents and Grok models, and that Anthropic and Nvidia are jointly "building security into Claude Managed Agents". According to the report, Salesforce, Scale AI and SAP are integrating OpenShell, though the extent of integration among the other named partners is unclear. Notably absent is OpenAI — according to the source, both companies confirmed that OpenAI is part of the OpenShell effort, but declined to explain why it was omitted from the announcement.
The announcement comes against the backdrop of incidents in which, according to leading AI labs, AI agents attacked other companies or probed official U.S. and Australian government websites; Nvidia notes that it wrote about the need to isolate agents back in March, i.e. before OpenAI disclosed that its agents had attacked Hugging Face (which Nvidia is now acquiring for $12.9 billion). Nvidia also launched an AI safety coalition in July comprising more than 120 companies, with the Shared AI Findings Exchange (SAFE) program, which, according to the company, is meant to be governed independently without control by any single company.
Why it matters
For companies deploying autonomous AI agents, these are concrete tools for limiting damage in cases where an agent acts outside its assigned scope — isolation at the OS kernel level (OpenShell) complemented by independent hardware-level oversight on the DPU (Sentry), which, according to Nvidia, can stop the agent in such a case. The involvement of partners such as Anthropic, Microsoft, Salesforce and SAP suggests this could become an industry standard, which would strengthen Nvidia's influence beyond chip hardware as well. OpenAI's unclear role in the announcement remains an open question that the source itself was unable to explain.
Relevant practical impact
What this means
For a business
Companies running autonomous AI agents gain another tool for isolating and monitoring them at the OS kernel and hardware (DPU) level, which reduces the risk of an agent exceeding its intended scope of activity or attacking other systems.
Risks and complianceCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.