The short version

Red Hat and Nvidia launched OpenShell, an open‑source sandbox that blocks AI agents from leaking data or spending money. The system adds default‑deny networking and policy checks to keep agents under strict control.

What happened

At GTC Berlin, Red Hat and Nvidia announced OpenShell, a sandbox runtime that enforces deny‑by‑default networking and policy checks for autonomous AI agents.

OpenShell uses Linux kernel mechanisms, Landlock for filesystem isolation, seccomp BPF for system‑call filtering, and network namespaces for network isolation, to keep agents outside their own control, according to Tech Times.

The project is open‑source under Apache 2.0, built in Rust, and will be integrated into Red Hat’s AI platforms as a first‑class capability, said SiliconANGLE.

The background

SiliconANGLE says that prompt‑injection attacks and silent data exfiltration have made AI‑agent security a critical concern for enterprises.

A Deloitte survey of 3,235 leaders found only 21% of organisations have a mature governance model for agentic AI, and Gartner has named AI‑agent oversight a top cybersecurity trend for 2026, reported by Tech Times.

The UK’s National Cyber Security Centre warned that prompt injection may be an inherent flaw of large language models, suggesting architectural fixes are needed.

Why is kernel‑level enforcement important?
Because it keeps control outside the agent’s code, preventing a compromised agent from disabling its own safeguards.

An agent shouldn’t enforce its own boundaries

John Fanelli, vice president of enterprise software, Nvidia

What it means for owners and operators

Owners can now deploy multiple agents with confidence that secret credentials stay hidden from the agents themselves, reducing the risk of data leaks.

Operators gain a policy engine that can update network and inference rules without restarting agents, resulting in less downtime and lower operational overhead.

Executives can attribute AI‑related costs more accurately, as OpenShell separates model access from execution, allowing quota and spend limits to be enforced centrally.

Founders building new AI‑agent products will have a ready‑made security blueprint, speeding time‑to‑market while meeting emerging regulations such as the EU AI Act.

Our take

Analysis by Infercline · Second-order effects

OpenShell gives enterprises a reusable guardrail to run multiple AI agents without risking secret leaks or runaway spend.

What to do in your business tomorrow

  1. Set up monitoring to alert when an agent attempts a prohibited system call or network connection.
  2. Create a central inventory of all credentials used by agents and rotate them regularly to minimise exposure.
  3. Define clear policy templates that restrict agents to only the APIs and data they need for their tasks.

Related coverage

Sources

  • siliconangle.com: Red Hat and Nvidia’s OpenShell puts AI agents on a deny-by-default leash
  • Tech Times: Red Hat and NVIDIA Built AI Agent Security Controls Agents Cannot Override
  • VentureBeat: OpenClaw launches free enterprise control plane for persistent AI agents, backed by OpenAI, Red Hat and Nvidia
  • Business Wire: Red Hat AI Factory with NVIDIA Accelerates Agentic AI Security Capabilities Across the Open Hybrid Cloud
  • Markets Insider: NVIDIA Launches OpenShell Reference Architecture for Safe AI Agents