skills

Least-privilege permissions

• Limit credential scope: Generate separate NVIDIA API keys for each agent skill with only the minimum permissions needed (e.g., read-only access for RAG, simulation-only for robotics). Never share one master credential across multiple skills.
• Set spending caps: Configure hard limits on GPU resource budgets per agent per month so runaway jobs cannot drain the budget unexpectedly.
• Restrict data access: Ensure the RAG skill can only read from approved document folders, not your entire company file system or sensitive databases.
• Rotate credentials regularly: Change NVIDIA API keys at least quarterly, and immediately after any employee departure or suspected breach.

Network isolation

• Firewall agent requests: Route agent traffic through a corporate proxy or VPN that logs and filters outbound API calls to NVIDIA services; block direct connections from agent environments.
• Sandbox simulation outputs: Run robotics and CUDA workflows in isolated compute containers that cannot execute arbitrary commands or access the host system.
• Segment RAG indexes: Store RAG training data on separate, restricted network drives with encryption at rest; never mix proprietary and public data in the same index.

Monitoring

• Log all API calls: Record who ran which skill, when, with what parameters, and the result—use NVIDIA's audit logs plus your own agent activity logs for full traceability.
• Alert on cost spikes: Set up automated notifications if GPU spending suddenly increases by 50% or more in a single day.
• Track credential usage: Monitor when and from which IPs each NVIDIA API key is being used; flag any unexpected locations or devices.
• Review RAG access patterns: Regularly audit which documents the RAG skill has retrieved to catch if sensitive files are being pulled into workflows unexpectedly.

Red flags to watch for

• Agents requesting sensitive files: If an agent tries to add employee records, financial data, or proprietary source code to a RAG index, block it immediately.
• Unusual compute spikes: GPU workloads that suddenly run for hours at 100% load without corresponding user requests may indicate a compromised skill or prompt injection attack.
• API key appearing in logs or error messages: If a credential is accidentally printed in agent output or error messages, treat it as compromised and rotate it at once.
• Agents generating code to run locally: If a skill returns executable commands or scripts rather than results, suspect prompt injection; do not run the code without manual security review.