OpenClaw Agent Compute

• Outcome: Review
• Latest audit: May 28, 2026, 9:49 AM
• Version:
  0.1.7

Overview

This looks like a legitimate compute-gateway client, but it gives an agent broad remote command and artifact control without enough safety scoping in the docs.

Install only if you trust and control the Compute Gateway and intend to let an agent run commands there. Use a sandboxed gateway, least-privilege revocable API keys, explicit approval for exec/delete/destroy actions, gateway-side command and path restrictions, logging, usage limits, and a pinned container image for the starter kit.

SkillSpector

By NVIDIA

Vulnerability Patterns

• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands
• Privilege Escalation: Excessive Permissions, Sudo/Root Execution, Credential Access
• Supply Chain: Unpinned Dependencies, External Script Fetching, Obfuscated Code
• Excessive Agency: Unrestricted Tool Access, Autonomous Decision Making, Scope Creep

Findings (1)

Missing User Warnings

Medium

• Confidence: 93% confidence
• Finding: The markdown advertises remote command execution (`POST /v1/exec`) and destructive operations such as artifact deletion and session destruction without any visible user-facing warning about execution risk, data loss, or cost implications. In an agent-friendly public skill, missing safety guidance materially increases the chance that users or upstream agents invoke dangerous actions without understanding their impact.

VirusTotal

64/64 vendors flagged this skill as clean.