agent-creator

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 1.00). The skill explicitly instructs agents to fetch and read untrusted, user-controlled web content and repos (e.g., the "browser-tester" example: "When invoked with a URL... Navigate to the target page" and the "repo-analyzer" pattern: "Clone the repo" in SKILL.md / references/patterns.md), so third‑party pages and public repositories are ingested and used to drive agent actions.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The agent's mcpServers frontmatter runs "npx -y @playwright/mcp@latest" (i.e., fetching and executing the @playwright/mcp package from npm at runtime), which clearly fetches remote code and executes it as a required dependency for the browser-tester agent.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (high risk: 0.70). The prompt explicitly documents and encourages use of powerful file-editing and shell tools plus permissionMode options like "bypassPermissions" and "acceptEdits" and hook commands that run arbitrary shell scripts, which enable agents to modify files and skip permission checks (a clear avenue to compromise machine state even though it doesn't explicitly call sudo or create users).