Back to skill

Security audit

Wan 3.0 Video Prompt Architect

Security checks for vulnerabilities and agentic risk

Overview

The skill is mainly a prompt-writing guide, but its optional setup tells users to run unpinned remote GitHub code locally, so it should be reviewed before installation.

Install the markdown skill only if you are comfortable using it as prompt-writing guidance. Treat the optional MCP setup separately: review the GitHub repository and dependency tree first, prefer a pinned commit or verified release, and run it with limited filesystem, environment, credential, and network access. Remove or avoid the MCP configuration if you only need prompt drafting.

Vulnerability Patterns
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Findings (1)

T03 · Remote Payload Retrieval and Execution

Error
Location
SKILL.md:62
Finding
Unpinned Remote MCP Source Is Downloaded and Executed<![CDATA[ ## Vulnerability Details **File Location**: `SKILL.md`, lines 62–72 **Vulnerability Type**: Execution of mutable remote code through `npx` **Risk Level**: High ### Vulnerable Code ```bash openclaw mcp add wan3-video \ --command npx \ --arg -y \ --arg github:gpt-img-2/wan3-video-prompt-mcp \ --include 'build_wan3_video_prompt,plan_wan3_shot_sequence,diagnose_wan3_video_prompt,get_wan3_video_resources' ``` Then run `openclaw mcp doctor wan3-video --probe`. If the installed release does not expose an `mcp` command group, add `npx -y github:gpt-img-2/wan3-video-prompt-mcp` as a Stdio server in MCP settings. ### Technical Analysis The installation instructions direct `npx` to retrieve and execute code from the mutable GitHub reference `github:gpt-img-2/wan3-video-prompt-mcp`. The reference is not pinned to a reviewed commit hash, immutable release, or exact package version, and no integrity digest or signature verification is required. Because the effective payload is resolved when the command runs, changes made to the remote repository after this Skill was audited can change the locally executed code without requiring any modification to `SKILL.md`. The `-y` option also suppresses the normal installation confirmation, reducing the opportunity for users to inspect the resolved dependency before execution. The audit found no evidence that the currently referenced repository is malicious. The vulnerability arises from trusting and executing a mutable remote payload without integrity controls. ### Attack Path 1. An attacker compromises the referenced GitHub account or repository, takes over a dependency used by the repository, or otherwise gains the ability to alter the remotely resolved source. 2. The attacker adds malicious package lifecycle logic, MCP server initialization code, or runtime code to the remote project. 3. A user follows the documented setup command or configures the fallback Stdio command. 4. `npx -y` downloads the current re ...[truncated 1154 chars]
Remediation
<![CDATA[ ## Remediation Suggestions 1. Pin the GitHub dependency to a reviewed immutable commit hash rather than the repository's current default branch. 2. Prefer an exact, trusted package release with a lockfile and verified integrity hash where the tooling supports it. 3. Remove `-y` from the documented command so installation requires explicit user confirmation. 4. Verify the source, release signature, dependency tree, and package lifecycle scripts before recommending execution. 5. Document the MCP server's required filesystem, environment-variable, credential, and network access. 6. Run the MCP server in a sandbox or container with: - A non-privileged account. - A read-only filesystem where practical. - No access to unrelated credentials or sensitive environment variables. - Restricted outbound network access. - Access only to explicitly required project directories. 7. Establish a controlled update process in which new upstream revisions are reviewed and the immutable pin is updated deliberately. 8. Provide the expected commit identifier and integrity verification steps directly in `SKILL.md` so users can confirm that the retrieved source matches the audited version. ]]>
Vulnerability Patterns
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (2)

Rp1

Medium
Category
MCP Rug Pull
Confidence
94% confidence
Finding
The skill instructs users to install and run an MCP server directly from a GitHub reference via `npx -y github:...` without pinning to an immutable version, tag, or commit. That creates a supply-chain risk: if the upstream repository is changed or compromised, users may execute attacker-controlled code locally when following the documented setup steps.

Skill Enumeration

Medium
Category
Agent Snooping
Content
- Pricing: https://wan-30.video/pricing
- Privacy policy: https://wan-30.video/privacy-policy
- Terms: https://wan-30.video/terms-of-service
- Raw Skill: https://wan-30.video/skills/wan-3-video-prompt-architect/SKILL.md
- MCP source: https://github.com/gpt-img-2/wan3-video-prompt-mcp
- ClawHub listing: https://clawhub.ai/gpt-img-2/skills/wan-3-video-prompt-architect
Confidence
80% confidence
Finding
Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Static analysis

No suspicious patterns detected.