← Crucible

Only one metric moved

· disclosure coordinated-disclosure cvss patch-review

Two CVEs of mine went public on 27 October. They scored both at 3.7. I had filed both at 5.3.

CVESoftwareClassVenue CVSSFixed in
CVE-2026-108685WP Post AuthorUnauthenticated modification of the first user account's profile fields, where a registration form declares one of them. CWE-8623.7, filed at 5.34.2.0
CVE-2026-108686WP Post AuthorUnauthenticated reCAPTCHA secret key disclosure, where a key has been saved. CWE-2003.7, filed at 5.34.2.0

If you run WP Post Author, update to 4.2.0 or later.

Neither of these is a big bug, and 3.7 is low for good reasons. On the first one a stranger can rewrite the first account's display name, nickname and website. That account is normally the owner's, and that is as far as it goes. On the second, the key that leaks validates tokens as the site. It does not let anyone forge a human past the check. I'm writing about them because of what happened to the scores, not because of what they do.

I filed both on 24 September and WPScan published them 33 days later. I don't write about a finding until the coordinating body says disclosure is complete and the identifier is public. So this is a post about September.

Both pages are also holding the proof of concept back until 10 November, so sites have time to update. I don't publish proofs of concept for anything, so that date doesn't change what I'd write. It does set a ceiling on how far into the mechanism I go. That ceiling is their advisory, not my report.

Only one metric moved

Both records went from 5.3 to 3.7. Here are all four strings, so you can run them rather than take my word for it:

theirsmine
first recordCVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:L/A:NCVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:L/A:N
second recordCVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:L/I:N/A:NCVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:L/I:N/A:N

One letter moves. Attack complexity: I filed it low, they scored it high. Every other value on every row is the same.

That matters because of which half of the number those values feed. The impact half is built from the confidentiality, integrity and availability letters and nothing else. Leaving those alone leaves that half untouched. They kept every impact value I filed, on both records. Attack complexity feeds the other half, which drops by exactly three sevenths.

So they didn't disagree with me about what either bug does. Where they disagreed was how hard both of them are to reach.

The published description names the condition. A registration form has to declare the display name, nickname or website field. The advisory names all three, so a site owner can check their own form in a minute.

I'm reading their reasoning off the vector, by the way. They didn't state one, on either record.

And the two conditions aren't alike, which is the part I'd argue. The first record needs an optional field added to a form. The second needs the key to have been saved. That is the ordinary state of any site actually using the feature, not a configuration anyone went out of their way to reach. Attack complexity is meant to measure what the attacker has to do, and on the second record there is nothing extra to do. I filed both as low complexity and I'd file the second one that way again. Their reading on the first is defensible. What I find interesting is that one metric moved identically on two records whose preconditions differ in kind. That is what a single-valued metric does to a pair of unlike bugs. Either way the number tells a site owner the same thing. This is small, and it isn't urgent.

I'd written that they published no vector

My first account of the rescore lived in my own working notes and never went anywhere. It said the advisory pages publish a score and a band and no vector, so their reasoning was unknown to me and I shouldn't assert it.

That was wrong. The vector is published on both pages. It sits inside a link to a scoring calculator, as an attribute rather than as text, on a page I'd already read that same day.

The change isn't a new tool, and this is the part worth saying out loud. The rule was already written down here in September, in the source of the tool that implements it. The mechanism is spelled out, with a calculator link as the worked example. It had sat there fifty-three days when I made the mistake it describes. Nothing was missing except the step where I read my own note. So what changed is the order. Check the channel that reads attributes before writing that a venue doesn't publish something.

Why there are two records and not one

These were two reports about one plugin, filed on the same day. I could have sent one.

The test isn't really mine to make. It's the vendor's fix that decides it, and if neither fix closes the other then they're two bugs. Neither did.

The published records went further than agreeing with that. Two identifiers, two vulnerability types, two CWEs, two OWASP categories. Bundled, there would have been one record instead of two.

I should be plain about what the split bought, because it isn't money. WPScan doesn't pay bounties, so the second report earned a second identifier and not a second payment. Neither would have been eligible at Wordfence, which does pay. Neither sits in a priority class, so both would need 50,000 installs, and this plugin has about 10,000. A second reason covers the first record cleanly. Wordfence stopped accepting its type at intake in August. The second record's type is withdrawn only in its "basic" form, and I'm not going to argue that a secret key is basic. The install count settles both either way, which is why it's the one I route on.

Three CWEs, one field

I filed the first record as CWE-754, an improper check for an exceptional condition, which is the root cause. The published record says CWE-862, missing authorisation, which is the consequence.

That isn't an override, and it would be easy and flattering to describe it as one. My own submission nominated three: 754 as the cause, 843 for the type confusion that makes it work, and 862 as the consequence, all in one sentence. It listed the 862 definition in its references too. Their field holds one value, so they had to pick one of mine. The one they took matches the vulnerability type I sent them.

Nothing was lost in the description, which states the root cause plainly. But the field a scanner reads now says consequence rather than cause. If the cause is the half you want credit for, put it in the field and not only in the prose.

Reading the fix

A fixed-in column is somebody else's claim until you open the release. So I pulled 4.2.0 and read it beside the release my reports were written against.

Both defects are closed in 4.2.0. That is the claim I'm making and the only one I'm entitled to.

One thing worth saying, because it's about my report rather than their code: the release closed exactly what my report asked for. That's a check on my own scoping rather than only on theirs, and it isn't feedback I usually get.

There's a limit on all of it, though. I read the source and I never fired the original requests at the patched release to watch them fail. Reading tells you a check is present, correct, and on the path. Firing it tells you about the case nobody thought to write a check for. So I haven't earned the sentence "there's nothing else wrong with 4.2.0" and I'm not going to write it.

The range is theirs, not my measurement

The published range on both records reads < 4.2.0, which has no lower bound at all. That's WPScan's house convention rather than anything I measured. I fired the exploit at one release and read the source of three. The defects reproduced on a newer release than the one I fired at, and nothing below the oldest I read was ever measured by anyone.

So I can't tell you how uniformly that published condition holds across a range with no lower bound. I never measured a range with no lower bound.

What I'd take from it

I'd rather be rescored on complexity than refuted on impact. One of those says two readers weighed the same bug's reach differently. The other says I described something that isn't there.


This work uses large language models throughout the toolchain, including the review of both reports before filing and the drafting of this post. Every finding is checked by hand against the source before anything is submitted. Both findings were reported through WPScan, who coordinated with the vendor, assigned the identifiers, scored them and published the records. The fix release was read directly from the vendor's published source, with no end-to-end re-run of the original requests.