Skip to main content
DataLeaked
Back to blog

Detector accuracy update: what changed and why

We corrected several measurements that were confidently wrong. Here is what changed in our font, WebRTC, cookie, and benchmark detection, and how to re-test.

Published: 2 min read

Why we audited

A full review of our detectors found cases where tools asserted things they did not measure: fonts reported as installed on every machine, WebRTC timeouts scored as clean, and cookies whose flags were never actually readable.

These were honesty failures, not feature gaps. The tools looked authoritative while being wrong, which is worse than showing nothing.

We fixed all of them and added regression tests so they cannot silently return.

What changed

Fonts: we no longer trust document.fonts.check alone, which returns true for any family. We now use a canvas measureText fallback that detects real rendering differences.

WebRTC: a timed-out probe no longer scores as 'no leak'. It is reported as partial/unknown and excluded from the score until a real result exists.

Cookies: we read flags through the correct API (CookieStore on window, not document) and label them unknown when the browser cannot expose them.

Trackers: the freshness guard that deleted tracker data on every score page load was fixed, so observed third-party services now feed the score.

Benchmarks: GPU benchmark now has a timeout and cancellation so a backgrounded tab cannot hang the button forever.

  • Fonts: canvas measureText fallback
  • WebRTC: timeout no longer scores as clean
  • Cookies: CookieStore on window, unknown when unreadable
  • Trackers: freshness guard fixed
  • GPU: timeout and cancellation added
  • Browser/OS detection: iOS, ChromeOS, and iOS third-party browsers fixed

What this means for your score

Some scores will drop, because previously inflated results are now honest. A lower score that reflects reality is more useful than a high one built on guesses.

Re-run your analysis after the update to get a current, accurate score. Use the before/after comparison to see the difference.

Factors that cannot be measured are now labeled unknown and excluded from coverage rather than scored as perfect.

How to re-test

Visit the privacy score page and run the full assessment. The guided flow walks you through the four required analyses.

Check the methodology page for the full list of what is measured, what is heuristic, and what is unavailable.

If you spot a result that still looks wrong, use the contact page to report it. We treat accuracy reports as high priority.

What we will not do

We will not inflate scores to look good. We will not guess when the browser cannot tell us. We will not treat an unavailable input as a perfect one.

These rules are documented in our methodology and enforced by tests.