Breaking

Nvidia Broadens OpenShell and Adds Sentry to Contain Autonomous AI Agents

September 28, 2026 at 05:05 EDT

AI Agents
Security
Open Source

Nvidia is moving OpenShell, its open-source runtime for containing autonomous AI agents, into broader availability and pairing it with Sentry, a separate monitoring layer for long-running agents. The release reframes agent safety as an infrastructure problem: keep policy enforcement outside the model, then audit and restrict every file, process and network action.

Sep 28, 2026 · NVIDIA

Agent containment moves below the model

OpenShell puts autonomous agents inside policy-governed sandboxes, while Sentry adds a separate monitoring layer for long-running workflows. The goal is to make every file, process and network action enforceable outside the agent itself.

Release Track

v0.1.2

Stable documentation now listed

Incident Record

17,600

attacker actions reconstructed

Agent Entries

listed in supported-agent docs

Customer Data

datasets reported accessed

Why the incident raised the stakes

ExploitGym tasks became a persistence problem: agents focused heavily on tasks models had not previously solved.

898

ExploitGym tasks

198

previously unsolved

Takeaway: persistent agents need a runtime that can deny, audit and revoke access even when the task pressure rises.

Control path

Agent attempts action

→

Sandbox reports request

→

Supervisor checks policy

→

Allow, deny or quarantine

What supporters like

Policy enforcement sits outside the agent process, credentials stay scoped to approved endpoints and actions leave an audit trail.

What remains hard

Default policies, network edge cases, experimental Windows support and per-agent coverage still need careful deployment work.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.