Skip to content

tc39/proposal-array-is-template-object

Folders and files

NameName
Last commit message
Last commit date
May 30, 2024
May 30, 2024
May 30, 2024
Mar 27, 2019
May 30, 2024
Mar 27, 2019
Mar 27, 2019
Mar 27, 2019
May 30, 2024
Oct 20, 2024
May 30, 2024

Repository files navigation

Reflect.isTemplateObject (stage 2)

Authors: @mikesamuel, @koto Champions: @littledan, @ljharb Reviewers: @erights, @jridgewell

Provides a way for template tag functions to tell whether they were called with a template string bundle.

Table of Contents

Use cases & Prior Discussions

Distinguishing strings from a trusted developer from strings that may be attacker controlled

Issue WICG/trusted-types#96 describes a scenario where a template tag assumes that the literal strings were authored by a trusted developer but that the interpolated values may not be.

result = sensitiveOperation`trusted0 ${ untrusted } trusted1`
// Authored by dev          ^^^^^^^^                ^^^^^^^^
// May come from outside                ^^^^^^^^^

This proposal would provide enough context to warn or error out when this is not the case.

function (trustedStrings, ...untrustedArguments) {
  if (Reflect.isTemplateObject(trustedStrings)
      // instanceof provides a same-Realm guarantee for early frozen objects.
      && trustedStrings instanceof Array) {
    // Proceed knowing that trustedStrings come from
    // the JavaScript module's authors.
  } else {
    // Do not trust trustedStrings
  }
}

This assumes that an attacker cannot get a string to eval or new Function as in

const attackerControlledString = '((x) => x)`evil string`';

// Naive code
let x = eval(attackerControlledString)

console.log(Reflect.isTemplateObject(x));

Many other security assumptions break if an attacker can execute arbitrary code, so this check is still useful.

An Example

Here's an example of how isTemplateObject lets a tag function wisely use a sensitive operation, namely Create a Trusted Type. The sensitive operation is not directly accessible to the tag function's callers since it's in a local scope. This assumes that TT's first-come-first-serve name restrictions solve provisioning, letting only authorized callers access the sensitive operation.

const { Array, Reflect, TypeError } = globalThis;
const { createPolicy } = trustedTypes;
const { isTemplateObject } = Reflect;
const { error: consoleErr } = console;

/**
 * A tag function that produces *TrustedHTML* or null if the
 * policy name "trustedHTMLTagFunction" is not available.
 */
export trustedHTML = (() => {
  // We use TrustedType's first-come-first-serve policy name restrictions
  // to provision this scope with sensitiveOperation.
  const policyName = 'trustedHTMLTagFunction';
  let policy;
  try {
    policy = createPolicy(
        'trustedHTMLTagFunction',
        { createHTML(s) { return s } }
    );
  } catch (ex) {
    consoleErr(`${policyName} is not an allowed trustedTypes policy name`);
    return null;
  }

  // This is the sensitive operation.
  const { createHTML } = policy;

  // This tag function uses isTemplateObject to reject strings that
  // do not appear in user code in the same realm.
  //
  // With a reliable isTemplateObject check, the attack surface is
  // <= |set of template applications in trusted code|.
  //
  // That set is finite.
  //
  // Without a reliable isTemplateObject check, the attack surface is
  // <= |set of attacker controlled strings|.  That is, in practice,
  // unbounded.
  //
  // This assumes no attacker has eval.
  const trustedHTMLTagFunction = (strings) => {
    if (isTemplateObject(strings) && strings instanceof Array) {
      return createHTML(strings.raw[0]);
    }
    throw new TypeError("Expected template object");
  };

  // With the check it's safe to export this tag function that closes
  // over a sensitive operation to anyone.
  return trustedHTMLTagFunction;
})()

Without isArrayTemplate, this can be bypassed:

// A naive, but non-malicious function.
function f(x) {
  // People trust trustedHTMLTagFunction.
  // Our HTML is trustworthy because <bad argument> so we'll just
  // piggyback off that by using a value that looks like a template object.
  // What could possibly go wrong?
  const s = dodgyMarkdownToHTMLConverter(x);
  const pseudoTemplateObject = [s];
  pseudoTemplateObject.raw = Object.freeze([s]);
  return trustedHTML(Object.freeze(pseudoTemplateObject));
}

// An attacker controlled string reaches f().
const payload = '<img onerror=alert(document.origin) src=x>';
console.log(`f(${ JSON.stringify(payload) }) = ${ f(payload) }`);

The threat model here involves three actors:

  • A team of first-party developers (in conjunction with security specialists) decides to trust the tag function.
  • A malicious attacker controls a string in the variable payload.
  • Non-malicious but confusable third-party library tries to provide a higher level of service by forging a template object. It assumes its clients are comfortable with trusting dodgyMarkdownToHTMLConverter to produce HTML for the current origin.

We've addressed this threat model when the first-party developers can be less tolerant of risk than the most risk tolerant third party dependency w.r.t. HTML injection.

This simple implementation doesn't deal with interpolations. A more thorough implementation could do contextual autoescaping.

What this is not

This is not an attempt to determine whether the current function was called as a template literal. See the linked issue as to why that is untenable. Especially the discussion around threat models, eval, and tail-call optimizations that weighed against alternate approaches.

Possible Spec Language

You can browse the ecmarkup output or browse the source.

Tests

The test262 draft tests which would be added under test/built-ins/Reflect

Related Work

If the literals proposal were to advance, this proposal would be unnecessary since they both cover the use cases from this document.