setup-sandcastle

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

• Third-party content exposure detected (high risk: 1.00). The REFERENCE.md headless templates explicitly call out running gh issue list (e.g., execSync('gh issue list ... --json number,title,body')) and pass ISSUE_TITLE/ISSUE_BODY into agent prompt files (implement.md), meaning the agent ingests untrusted, user-generated GitHub issue content that can influence its actions.

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

• Potentially malicious external URL detected (high risk: 0.90). The onSandboxReady hooks run bunx commands that fetch and install remote skill repos (e.g., malinskibeniamin/skills/frontend-starter-kit and malinskibeniamin/skills/development-lifecycle) at runtime, and those fetched skills are then used as active prompt/instruction modules for agents, so they directly control agent behavior and execute code in the sandbox.