Filesystem
Files within Reach.
Filesystem policies define which paths an agent can read or change. Linux Landlock enforces that boundary below the agent.
How it works
OpenShellPowerful Agents. Boundaries Beyond the Model.
Intelligence needs room to work. NVIDIA OpenShell helps us give it a controlled environment, with permissions enforced outside the model.
User
Globster Control Plane
OpenShell RuntimeSecrets VaultCredentials kept outside agent context
Isolated Sandbox AAgentTools · Plugins · Skills
Persistent Workspace
Isolated Sandbox BAgentTools · Plugins · Skills
Persistent Workspace
File & Process Controls · Outbound Network Policy
Permitted outbound paths
Public Internet
WHAT IS OPENSHELL?
A Runtime Built for Agent Access.
NVIDIA OpenShell is an open-source runtime that places policy-enforced boundaries around AI agents. It controls access to files, processes, networks, and provider credentials outside the model itself.
Read NVIDIA’s overviewNetwork
Connections with Permission.
Outbound connections pass through policy checks. Destinations and executables need matching rules; unlisted connections are denied.
How it worksProcess
Restricted Execution.
An unprivileged process and syscall restrictions reduce what agent code can do to its surrounding environment.
How it worksCredentials
Credentials Outside the Model.
Supported provider connections use proxy-managed credentials. The agent can use an authorized service without receiving its raw provider key.
How it worksWHY GLOBSTER CHOSE IT
Control That Travels With the Work.
Globster uses NVIDIA OpenShell as the secure runtime layer for hosted agents. We combine that runtime boundary with organization controls, connected tools, credential handling, and human approvals.
Explore Globster securityEnforcement Outside the Agent.
A model can make an unexpected decision. The runtime evaluates access separately, so instructions in a conversation are only one part of the safety design.
Policies Matched to the Work.
Globster translates network rules into OpenShell enforcement, including destination, executable, and configured HTTP method and path restrictions.
Layers We Can Inspect.
An open runtime gives our team an architecture to examine and controls to validate, alongside Globster’s identity, access, and approval systems.
RECENT INCIDENTS, PRACTICAL LESSONS
Take the Risk Seriously.
Recent OpenAI disclosures show why agent security needs several layers. These reports concern internal research or deployment environments; they describe different failure modes.
DNS Can Be a Route Out.
On September 20, 2026, an OpenAI research agent reached an external chatbot through insufficient DNS filtering. The report describes a network restriction bypass, not a demonstrated host-kernel escape.
A Readable Key Can Be Leaked.
OpenAI disclosed that an internal model published a researcher’s GitHub token while pursuing a task, despite instructions against the behavior. The incident occurred on May 27, 2026.
Reachable Services Carry Risk.
OpenAI’s Hugging Face investigation describes a platform compromise driven primarily by an internal research model. Access to an external service can create consequences beyond the sandbox.
OpenShell helps enforce configured boundaries. It does not establish that a particular deployment is immune to these incidents. Allowed services, DNS paths, software updates, and policy configuration still need validation.
The Questions Worth Asking.
Understand the protection, and the boundaries of the claim.
Still have a question? Talk to our team
Let’s Talk About Your Boundaries.
Walk through your tools, data, and security requirements with our team.
NVIDIA’s agent security guidance