cd /news/ai-safety/how-one-web-page-could-plant-script-… · home › topics › ai-safety › article
[ARTICLE · art-170706] src=theainewsreport.com ↗ pub= topic=ai-safety verified=true sentiment=↓ negative

How one web page could plant script on every site through uBlock Origin: the element picker bug GPT-5.6 Sol found, step by step

A universal cross-site scripting flaw in uBlock Origin's Firefox MV2 build let a malicious web page drive the extension's element picker and add a trusted filter that could inject script into any site, according to researcher Smitty van Bodegom's November 15, 2026 report. The advisory GHSA-cmqw-9r2w-6wv2, rated High, lists uBlock Origin 1.75.2 and lower as affected and 1.75.4 as patched, with four commits shipped between November 2 and November 8; the latest stable release is 1.76.4, published November 15. Van Bodegom said GPT-5.6 Sol did most of the work of finding the bug, which he verified himself, and that Claude wrote most of the proof-of-concept code.

by read6 min views2 publishedNov 16, 2026
How one web page could plant script on every site through uBlock Origin: the element picker bug GPT-5.6 Sol found, step by step
Image: Theainewsreport (auto-discovered)

A researcher used GPT-5.6 Sol to find a one-click bug that let a website add trusted filters to uBlock Origin, which can inject script into any site. We walk through the chain and match each step to the commit that closed it.

This explains reporting by

Smitty van Bodegom: Universal XSS in uBlock Origin from website-driven element picker (November 15, 2026). Read the original first:

https://iter.ca/post/ubo-uxss/

In one minute #

  • uBlock Origin's Firefox MV2 build let a malicious page drive its element picker and add any trusted filter after one click. Trusted filters can inject script, so the page could run code on any other site.
  • The researcher says GPT-5.6 Sol did most of the work of finding the bug. He verified it himself, and Claude wrote most of the proof-of-concept code.
  • The advisory (GHSA-cmqw-9r2w-6wv2, rated High) lists 1.75.2 and lower as affected and 1.75.4 as patched. Four commits shipped between November 2 and November 8.
  • uBO Lite (MV3) had a similar but less serious issue, per the researcher.
  • Admins: make sure managed browsers run 1.76.x. The latest stable release is 1.76.4, published November 15.

What the element picker is #

uBlock Origin has a tool you open from the right-click menu to point at something on a page and hide it. It writes a filter rule for you, without leaving the page.

The picker is a small page inside the extension, loaded into an iframe at an address like moz-extension://<random ID>/web_accessible_resources/epicker-ui.html. That page can tell the extension's background script to add filters. Filters added through the picker are treated as trusted, at least until the filter lists are compiled again.

Trusted filters matter because some of them can run scriptlets, small pieces of JavaScript, on the sites they match. A filter that is both trusted and attacker-written is a way to run attacker code on any site.

The chain, step by step #

  • Step 1, learn the extension's random ID. On Firefox the ID is random per install. A default filter swaps an ad script request for a harmless stand-in file served from the extension's own address. The attacking page hooks the browser's XMLHttpRequest, triggers that filter, and reads the moz-extension address, which carries the ID.
  • Step 2, load the picker. With the ID, the page sets an iframe's address to the picker page.
  • Step 3, drive the picker by message. The page opens a message channel to the iframe and sends the same messages the real extension sends to start the picker and fill in a filter. One field, opt, is meant to hold a request type such as image. Because it was not checked, a newline in it adds a second line, and that line becomes its own filter, a trusted script filter for another domain.
  • Step 4, get one click. The iframe can be invisible (opacity 0) and placed under something the user will click, so the click lands on the picker's Create button.
// what the picker expects in opt:      "image"
// what the page sent instead:          "image\n<second filter for another site>"
// result: the second line is saved as its own trusted filter

Which fix closes which step #

The GitHub advisory lists the commits in order. Matching them to the chain:

  • 1.75.2, November 2, commit bf69cf7, 'Avoid use of potentially tainted XHR methods in scriptlets'. This targets step 1. The same day, the reporter showed the ID could still leak through a responseText getter, so this fix alone was not enough.
  • 1.75.4, November 3, commit 9859477, 'Launch element picker only with a verified secret token'. This breaks step 3: the real picker now gets a secret from the extension, so a page cannot drive it. The advisory lists 1.75.4 as the patched version.
  • 1.76.0, November 8, commit b1f7d27, 'Validate about to be created filter in element picker'. This closes the newline trick in the opt field.
  • 1.76.0, November 8, commit 150df12, 'Prevent war:-related filters for the time being'. This removes the kind of filter that leaked the ID in step 1.

The order shows a good fix pattern: block the leak, then block the control channel, then validate the input, then remove the risky feature. Any one of the later three stops this attack; together they cover variants.

The AI part #

The researcher says he suspected filter injection could lead to cross-site scripting, asked GPT-5.6 Sol to look, and it found the issue. He then checked and verified it. He says Claude wrote most of the proof-of-concept code, and that he wrote the blog post himself.

He reported it by email on November 1. The first fix shipped the next day, and the details went public on November 15, after a stable release was out. The disclosure came in the right order.

It is one more example of a model doing the search work on a mature, widely reviewed codebase, with a human confirming the result. ZDNET reported the same weekend on a pen-testing firm's test in which a model found a V8 bug for $4.24.

Who is affected #

Case Status
uBlock Origin for Firefox (MV2), 1.75.2 and lower Affected, per the advisory. Update.
uBlock Origin 1.75.4 and later Patched, per the advisory. 1.76.0 adds filter validation.
uBO Lite (MV3) A similar, less serious issue, per the researcher. The advisory does not list separate versions.
uBlock Origin on Chromium MV2 The report says the same hook could read the extension's long secret there too; the advisory does not list a separate version.

What to do #

  • Check the uBlock Origin version on every managed browser and push 1.76.x (1.76.4 is current).
  • If you deploy the extension by policy, confirm automatic updates are not blocked.
  • Treat browser extensions as software you patch, with an owner and a version check, like any other client app.

What is still unknown #

  • Whether the bug was ever used in an attack. Neither the post nor the advisory says it was.
  • No CVE ID has been assigned; the advisory says 'No known CVE'.
  • The details of the uBO Lite issue. The researcher calls it similar but less problematic and gives no steps.
  • The researcher notes there are several ways the extension ID could leak, not just the one filter he used.

Sources #

AI News Report· every headline, every morning.

── more in #ai-safety 4 stories · sorted by recency
── more on @ublock origin 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/how-one-web-page-cou…] indexed:0 read:6min 2026-11-16 · —