Get unstuck

Hint 1: Permission is data outside the document text

The document body is untrusted content, even when it says to ignore filters.

Hint 2: Tenant membership is only the first boundary

Within a tenant, document groups still restrict access.

Reveal the reference patch and explanation
Fixedretrieval.mjsJavaScript
export function retrieve(user, documents, query) {
  if (!user) return [];
  return documents.filter(d => d.active === true && d.tenant === user.tenant
    && d.groups.some(g => user.groups.includes(g))
    && d.text.toLowerCase().includes(query.toLowerCase()))
    .map(d => d.id).sort();
}

For North/staff and query invoice, only n-public passes active=true, tenant equality and group overlap. n-payroll fails the group check; s-report fails tenant; n-retired fails active; n-injection fails group despite text that asks for broader access. South/staff receives only s-report. North/finance with INVOICE receives n-payroll because relevance is case-insensitive after authorization. Removing all groups returns no documents. These result ids stand in for the content that a real retriever would pass onward; no model is invoked.

Compare it with your own fix, then rerun the tests. The editor is not changed when you open this explanation.

Inspect the actual test harness

The browser and the lab download run the same cases, including cases that catch common wrong fixes, such as comparing with one fixture value instead of the real rule. The download includes the complete test adapter and cryptographic fixtures.

Download cases.mjs
How the browser sandbox runs your code

Your code runs in QuickJS 2025-09-13 (quickjs-emscripten 0.32.0) inside a Web Worker, isolated from the page, with a 32 MiB memory limit and a 2.5-second time limit. The test cases and the pass or fail verdict run outside that sandbox, so changing built-ins such as Array.prototype.push cannot change the result. The lab download runs the same cases under Node.js, where your code runs directly.

The same as in current Node.js:

  • ES2025 syntax and built-ins, including Set methods such as intersection, iterator helpers, Promise.try, RegExp.escape and the RegExp v flag

Added to the sandbox so that it behaves like Node.js:

  • structuredClone, Array.fromAsync, queueMicrotask, atob and btoa
  • URL and URLSearchParams, parsed by your browser’s own URL implementation
  • TextEncoder and TextDecoder for UTF-8
  • Intl.Collator and a locale-aware String.prototype.localeCompare
  • console.log and related methods, whose output appears with your results

Not available in the browser sandbox:

  • Timers (setTimeout, setInterval), fetch and other network or storage APIs; the exercises are deterministic and offline
  • Node.js modules and globals: import, require, process and Buffer
  • Web Crypto (crypto.getRandomValues, crypto.subtle); each workshop supplies the cryptographic operations it needs
  • Intl other than Collator: toLocaleString and toLocaleDateString ignore the locale, so numbers have no digit grouping and dates use one fixed format in the device’s time zone
  • using declarations, Symbol.dispose and DisposableStack (explicit resource management)
  • Uint8Array.fromBase64 and toBase64, JSON.rawJSON, Atomics and Error.captureStackTrace

Use the lesson in your own system

The exercise uses exact substring relevance and supplied authorization data. It does not run embeddings or a model. A real vector store must enforce equivalent permissions, scoped caches and revocation behavior across indexes, rerankers, citations and logs.

Your next task: Which cache keys and indexing updates carry your authorization scope and permission revision?

OWASP GenAI security project

Related guide: Keep retrieval inside the user’s data boundary