vercel-sandbox

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

• Third-party content exposure detected (high risk: 1.00). The skill's workflows (e.g., screenshotUrl, snapshotUrl, fillAndSubmitForm, and the Cron example in SKILL.md) explicitly open and snapshot arbitrary URLs with agent-browser and parse/access the page (accessibility snapshots, element refs, fill/click), meaning the agent ingests untrusted third‑party web content that can influence subsequent actions.

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

• Potentially malicious external URL detected (high risk: 0.80). This skill explicitly runs package installs and npx commands at runtime (pnpm add @vercel/sandbox, npm install -g agent-browser, npx agent-browser ...) which fetch and execute remote code from the npm registry (e.g. https://registry.npmjs.org/@vercel/sandbox and https://registry.npmjs.org/agent-browser), so required external content is fetched and executed during runtime.

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

• Attempt to modify system services in skill instructions detected (high risk: 1.00). The prompt instructs the agent to run sudo package-install and ldconfig commands (and to create sandbox snapshots that modify VM system state), i.e. it directs privileged system-level changes requiring sudo, so it pushes the agent to modify the machine state.