Building More Secure Autonomous AI Agents with NVIDIA NemoClaw

An autonomous agent can do more than generate text. It can execute commands, modify files and call external APIs. Controls that inspect model inputs and outputs do not, by themselves, establish boundaries around those actions.

Runtime controls restrict the resources and operations available to the agent. The session explains why teams should enforce these boundaries outside the agent, so protection does not depend solely on the model following instructions.

OpenShell combines sandboxing with policies that govern filesystem access, process behavior and outbound network requests. The architecture discussed in the webinar includes Landlock for filesystem restrictions and seccomp for filtering system calls.

Network policies define which external services an agent can reach and how it can interact with them. These controls help limit the consequences of unwanted behavior, but their effectiveness depends on the permissions granted. Broad access can still allow harmful actions within those permissions.

The setup separates the machine running the agent from the service running model inference. Nebius AI Cloud hosts the agent environment, while Nebius Token Factory provides an OpenAI-compatible inference endpoint.

The walkthrough identifies the required provider settings: the API base URL, the exact model identifier and an API key. It also introduces a deployment skill that a coding agent can use to help configure the environment. Cloud access and inference credentials must be prepared before that workflow can run.

Web search gives an agent access to information beyond its model’s training data. The webinar uses Tavily as the search service and explains how to configure its API key during OpenClaw onboarding.

Users can interact with OpenClaw through a terminal, its dashboard or a connected messaging service. The recording shows a Slack bot example, while the dashboard provides access to settings, sessions, skills and scheduled jobs. The full Slack application setup is outside the scope of this session.

Skills and plugins extend what an agent can do, including connecting to services and operating cloud infrastructure. They also introduce instructions and code that may access sensitive resources.

The speakers recommend reviewing where extensions come from and what access they require. They caution that community repositories can contain malicious skills, even when scanning is available. The deployment helpers shared during the session are presenter-maintained examples, so teams should review them before using them in their own environment.

Start with least privilege: give the agent only the files, services and operations needed for its task. Protect dashboard access with a strong authentication token, handle API credentials carefully and review permissions when adding integrations.

The session also recommends monitoring usage and configuring available spending limits. Sandboxing is one part of operating an agent responsibly; teams still need to manage credentials, review extensions and understand the access their policies permit.