Nvidia has announced an Open Agent Safety Platform that combines OpenShell 0.1.0, a kernel-enforced runtime for isolating AI agents, with Nvidia Sentry, a watchdog service running on the company’s BlueField-4 data processing units. The platform is designed to control agents’ access to networks, tools and systems outside the model itself.
Safety outside the model
The launch follows reports from OpenAI, Anthropic, Meta and Google describing agents that reached real systems from test environments. Nvidia’s position is that model training and alignment cannot reliably determine everything a probabilistic agent may access, making external enforcement necessary for high-risk deployments and evaluations.
OpenShell places each agent in a kernel-isolated sandbox. Network access is blocked unless it passes through a supervisor outside the workload, keeping security controls beyond the agent’s reach. Nvidia announced the runtime at GTC in March and is releasing version 0.1.0 as part of this platform.
A prover checks combined permissions
The main new OpenShell feature is a policy prover. It uses deterministic mathematical reasoning to check whether permissions remain within the boundary set by an operator, including combinations created when multiple agents communicate or delegate work. That is intended to expose indirect paths that separate permission checks could miss.
Nvidia says its tests involved agents with reduced safeguards repeatedly trying to persuade an AI reviewer to grant write access to a protected repository. According to the company, the prover supplied evidence about the request and no protected writes occurred. This is Nvidia’s test claim, not an independent evaluation.
Sentry adds optional hardware control
Sentry runs in a separate trust domain on BlueField-4 and monitors traffic between agents and their model endpoints. Nvidia says it can quarantine an agent in milliseconds and cut it off at the network level. The DPU is optional: Nvidia says OpenShell on CPUs can provide strict access control, while Sentry is aimed particularly at frontier-model evaluation and red-team environments.
Sentry is not open source, although Nvidia says it will provide open APIs. OpenShell can also work with other network-enforcement hardware. Anthropic, SpaceXAI, Salesforce and SAP are listed as users or contributors, while OpenAI, Google and AWS are absent from the announced partner list.
AI-generated editorial illustration; not a documentary photograph.
Source: Nvidia launches Open Agent Safety Platform to lock down rogue AI agents
SIGNAL TITANS
Leave a Reply