DEV Community

OnePassApps
OnePassApps

Posted on

A Collision in Our Own Per-Request Key: One of Our URLs Returned the Same Bytes Twice in 30 Fetches

I work on OnePassApps - the team behind app.onepassapps.com - so this is our own site and our own build. Every number below was measured on it today, 12 November 2026, signed out, over plain HTTPS, with no browser and no JavaScript.

On 10 November I published a post here saying that one of our URLs never returns the same bytes twice. I need to correct that. Today I fetched one of our pages thirty times in a row and two of those thirty responses were byte-for-byte identical.

The 10 November post is the canonical write-up on what moves between two fetches. This post is about the exception I missed.

The part that still holds

The mechanism is exactly as described then. I fetched all 56 URLs in our sitemap twice today. All 56 came back with the same byte length on the second fetch. Between the two copies of each page, 4,914 bytes differed across 404 runs - and every one of those bytes sat inside a run of long hexadecimal characters. None sat outside. So the differences live in one place, and that place is the encoded copy of our contact address.

What I got wrong was the word "never".

The key is one byte wide, so it repeats

Each of our pages carries our support address in Cloudflare's email-protection encoding: 42 hexadecimal characters, which is 21 bytes. The first byte is a key. The remaining 20 bytes are the address with every character XORed against that key - and 20 characters is exactly the length of the address we print in the footer.

There are 256 possible keys. That is the whole supply. It is not a fresh value per response, it is a draw from a set of 256.

The collision

I fetched one of our guide pages thirty times, a moment apart. The thirty responses produced 28 distinct SHA-256 digests - two digests each appeared twice. The key byte repeated twice across the thirty fetches, and in both cases the response that repeated a key was byte-for-byte identical to the earlier one with that same key.

Thirty fetches give 435 pairs, and each pair has roughly a one-in-256 chance of drawing the same key, so a key spread evenly over those 256 values would be expected to collide about 1.7 times. I saw 2. The page in question stayed exactly 28,749 bytes long across all thirty fetches.

Why it matters to anyone measuring a site

A hash comparison between two responses can tell you the bytes differ. It cannot tell you the page was rebuilt - but the reverse case matters: two responses with the same digest may be two draws of the same key rather than a page that stayed still. Any pipeline watching our pages for changes will occasionally see a flat line, and it will not be able to tell a quiet page from a repeated key.

So if you record a page size or a page hash once and quote it later, you are quoting one sample from a range - and occasionally the range collapses to a single value.

The page that did not move at all

In the 56-page sweep, 55 pages came back with a different digest and one - the same guide page - came back byte-for-byte identical. That is the collision above, caught on a double-fetch by chance.

The one page I did not explain

One page in the sweep changes far more than the rest: 1,449 bytes across 96 runs, against 42 or 84 bytes on 45 of the 56 pages. Those bytes are still all inside long hexadecimal runs, so they come from the same mechanism, but I did not analyse them and I am not claiming what they are.

What this does not say

One build, one day, one network, signed out, one GET per fetch, plain HTTPS, no browser, no JavaScript, no cookies. I did not test from a second network, did not test a signed-in request, and I make no claim about how any cache or browser treats these responses. The 10 November post measured a double-fetch of the same 56 pages and found 0 that were byte-identical; a double-fetch is a small sample, and today one of them was.

You can see the site at https://app.onepassapps.com?utm_source=devto&utm_medium=community&utm_campaign=key-collision&utm_content=app-link

And the guide page where the collision showed up is at https://app.onepassapps.com/guides/paperkit?utm_source=devto&utm_medium=community&utm_campaign=key-collision&utm_content=paperkit-link

The 10 November post is the canonical write-up on the per-request encoding: The Same URL Never Returns the Same Bytes: 56 Pages of app.onepassapps.com and One Per-Request Key

I work on OnePassApps - the team behind app.onepassapps.com. Corrections are better than praise.

Top comments (0)