178
organisations with an infected Optimizely snippet, per Report URI
68
affected organisations Report URI tested that had no CSP at all
24
tested organisations where a restrictive script-src blocked the payload
$7,040
received by the attacker's wallets as of 19:25 UTC on 16 Nov 2026
A tag vendor's dashboard is a production deploy path
The code that ran on Zoom, the BBC and CNN on 16 Nov 2026 did not come through a pull request. It came through a text box. Optimizely Web lets a customer paste custom code into a field called Project JavaScript, and that code is then served inside the customer's snippet from cdn.optimizely.com and, in Report URI's words, "runs on every page". Someone other than the customer wrote to that field across many accounts at once. Report URI found 188 infected snippets belonging to 178 organisations.
That is the argument of this post. An experimentation or tag platform is a way to ship JavaScript to production without a build, a review or a release. Most organisations treat it as a marketing tool. An attacker treats it as the shortest path to every visitor's browser, and on 16 Nov that path was open.
The second half of the argument is about the control most sites believed they had. A Content Security Policy that lists the vendor's host as an allowed script source says one thing: code from that host may run. It says nothing about who wrote the code. The attacker's script arrived from the exact host the policy trusted, inside the exact file the site asked for. Report URI tested the affected organisations' own hosts and found that a restrictive policy blocked the payload for 24 organisations. For the rest, the browser did what it was told.
We do not know how the attacker got write access, and neither does anyone who has published. So the useful question for a defender is narrower and can be answered today: if a vendor's file turns hostile at four in the morning, what in your own estate notices, and what stops it?
What Report URI and Hudson Rock recovered
Three write-ups were published on 16 Nov 2026, and they agree on the core. Report URI fetched 1,611 distinct Optimizely snippets, found mostly through its own crawl data, and checked each one. Its peak sweep, between 19:24 and 19:59 UTC, found 186 infected, which is where the figure in early news posts came from. The final count was 188 snippets across 178 organisations, and Report URI says the real number is likely higher because it could only check the snippets it could find in time. Its headline names include Zoom, the BBC, CNN, Salesforce, GitLab, IKEA, HSBC, Equifax, Air France-KLM, Ledger and Peacock. Optimizely's own websites carried the code too.
Hudson Rock worked from recovered files. It pulled eight poisoned Optimizely assets directly, documented 21 named web properties with evidence, and is explicit that this is "a documented subset of the activity, not an incident-wide victim count". An independent researcher, Amirul Cyber, published the loader, a table of snippet paths by site, and an indicator list.
-
Where //
the Project JavaScript field of customer snippets on cdn.optimizely.com -
Scale //
188 snippets, 178 organisations, of 1,611 snippets checked (Report URI) -
First seen //
04:21 UTC, in the BBC's snippet (Report URI) -
Mostly removed //
one infected snippet left at 21:52 UTC (Report URI) -
Attacker host //
1688360[.]cc, plus webshell[.]lol and several .icu hosts -
Root cause //
not known; Optimizely has not commented
The timeline, all in UTC from Report URI, shows an operator working through the day. The last clean capture of the BBC's snippet was at 02:21. The first injection appeared at 04:21. The main payload file, track.js, was first seen on bbc.com at 08:28. At 13:03 the attacker swept stolen cryptocurrency out to exchanges. At 13:43 a card skimmer was added. Between 17:13 and 17:57 a second iPhone host was added and the Coruna frame was injected into every visit from iOS 18.6 or earlier. Clean-up ran from about 20:40, and by 21:52 every snippet Report URI tracked except one had been cleaned. Report URI does not say whether Optimizely or its customers did the cleaning.
This was also not the operator's first day. Report URI and Hudson Rock both describe an earlier wave from 13 Nov that put the same injector into scripts from three other providers: Kissmetrics, StorifyMe and Minute Media. Hudson Rock's earliest capture is a Kissmetrics response on a bank's site at 05:42 UTC on 13 Nov. Those sites are not in the 178.
What the code did, in four parts. It replaced cryptocurrency addresses on the page with the attacker's own across eight address families, including Bitcoin, Ethereum, TRON and Solana. It read card numbers, expiry dates and security codes out of form submissions and posted them to the attacker's host. It took the page address, the cookies that script can read, and a screenshot. On iPhones and iPads it loaded a second stage from separate hosts that routes by iOS version to pages associated with two known exploit kits, Coruna and DarkSword.
The limits matter as much. Report URI says plainly, "We cannot see how many cards or cookies reached" the attacker. Hudson Rock found no routine that reads local storage, notes that cookies marked HttpOnly are out of the script's reach, and says the code does not show a way past a payment form hosted in a cross-origin frame. On the phones, its test was a headless browser with an iPhone user agent, and its wording is exact: "These are delivery observations, not demonstrated iPhone compromises." The one hard number for harm is small. Report URI measured about $7,040 received by the attacker's wallets as of 19:25 UTC.
How one field in a snippet became a skimmer
Each stage here is ordinary. The damage comes from where the first stage sits.
Figure 01
From a dashboard field to the visitor's card
- 01
Write //
Attacker code lands in Project JavaScript.
- 02
Serve //
The vendor's CDN delivers it as your snippet.
- 03
Load //
It fetches a payload, or runs a built-in copy.
- 04
Take //
Addresses swapped; cards and cookies read.
- 05
Route //
iPhones are sent on to exploit kit pages.
Result //
One vendor field reached the visitors of 178 organisations.
Hudson Rock found the injection sitting inside the customer's Project JavaScript, next to the legitimate experiment code. Nothing else in the snippet was touched, which is why the file kept working and nobody's tests broke. Report URI rules out cache poisoning: "Optimizely's CDN served exactly what it was given."
The loader then did something that defeats the first instinct of most responders, which is to block the bad domain. It asked the attacker's host for the live payload. If that request failed, or had not finished after 2.5 seconds, it ran a full copy of the payload that was already inside the snippet. Hudson Rock recovered that built-in copy, 24,392 bytes of it, from the BBC and HSBC UAE specimens, and warns that domain blocks alone are not enough because of it. A firewall rule against the attacker's host changes which copy runs. It does not stop the code.
From there the payload behaves like any script on the page, because it is one. It can read what the visitor types, rewrite what the visitor sees, and watch what the page sends. Hudson Rock describes the address swapper rescanning the page every 50 milliseconds while it is visible, and the card module hooking form submissions and outgoing requests. The visitor's browser has no way to tell this code from the site's own.
Why allowing the vendor's host protected nobody
Report URI did the measurement that makes this incident worth more than its dollar figure. It crawled 256 of the affected organisations' own hosts, evaluated each policy against what the payload needs, and sorted the organisations into four outcomes.
What the site's own policy did, per Report URI
-
No CSP //
68 organisations. Everything ran. -
CSP allowed it all //
50 organisations. The payload ran and the stolen data left. -
Half stopped //
17 organisations. The payload ran, connect-src blocked the theft, addresses were still swapped. -
Blocked //
24 organisations, which Report URI puts at 15%.
Read the second row twice. Fifty organisations had written a policy, deployed it, and were no better off than the 68 with none. Report URI's explanation is short: what separated the 24 protected sites from the rest was a restrictive script-src, and allowing cdn.optimizely.com "was never going to help". A policy built by listing every host your tags load from is an inventory. It records what you already trust. It does not constrain it.
Two keywords did the damage. Report URI found that policies carrying 'strict-dynamic' or 'unsafe-eval' let the payload through. The reason is the built-in copy. When a policy blocked the remote payload, the loader fell back to the copy inside the snippet and ran it through eval. Report URI saw exactly that at HSBC, whose policy blocked the remote file and the data theft, and where "the inline copy still ran". A site without 'unsafe-eval' refuses that step. Report URI notes that Optimizely's own documentation tells customers to allow it.
The third row is the instructive one for anyone who takes payments. Those 17 organisations did not stop the code, but their connect-src did not list the attacker's host, so the browser refused the outgoing calls. MDN lists what connect-src governs: fetch, XMLHttpRequest and sendBeacon among them, which are the calls the card and cookie modules use. The cards stayed in the browser. The address swap still worked, because changing text on a page needs no network call at all.
So the policy that would have held has three properties. Its script-src is restrictive and does not allow eval. Its connect-src names the hosts the site really talks to and nothing else. And it reports, so that the first blocked request at 04:21 becomes an alert and not a line nobody reads. None of that is new advice. What is new is a same-day measurement of how few large organisations have it: 24 organisations blocked the payload and 135 ran it.
How to check your own site, and what we saw at 22:50 UTC
Start with what your visitors receive, not with what your dashboard shows. Fetch your own snippet from cdn.optimizely.com and read the Project JavaScript body in it. Report URI's advice is to look for the attacker's host, for atob( calls, and for any script you did not write, in every project and in each experiment's custom code. Hudson Rock adds an order of operations that incident responders will recognise: preserve the bytes, hashes, headers and project revisions first, then remove.
We ran that check ourselves from outside. At 22:50 UTC on 16 Nov we fetched thirteen of the snippets named in the public write-ups and searched each for the published indicators, in plain text and in Base64. None of the thirteen carried them. Their Last-Modified headers fell between 19:51 and 22:43 UTC, consistent with the clean-up window Report URI describes. That is one fetch from one network location, an hour after Report URI's last sweep. It shows those thirteen were clean when we looked. It does not show the incident is over.
One of the thirteen was odd. In the snippet that Amirul Cyber ties to an HSBC insurance site in Singapore, the entire Project JavaScript body was 24,200 repeats of a single character, with a Last-Modified time of 21:10 UTC. We cannot tell who wrote that or why. It reads like a blunt overwrite of the field, and if so it is worth knowing that a fix can replace the attacker's code without restoring yours. After any clean-up, compare the field with the last version you know you wrote.
For your own network, the attacker's hosts are worth a retrospective search. This only sees devices behind your proxy. It says nothing about your customers' browsers, which is where the harm was.
Figure 02
Requests to hosts tied to the Project JavaScript injection
Sigma
title: Requests to hosts tied to the Optimizely Project JavaScript injection
status: experimental
description: Proxy requests to the hosts published by Report URI and Hudson Rock on 16 Nov 2026
logsource:
category: proxy
detection:
selection:
cs-host|endswith:
- '1688360.cc'
- 'webshell.lol'
- '5vobauua3ojpklx.icu'
- 'eil786npto9u5bq.icu'
- 'l7bcsnwc7nzr3uo.icu'
- 'je4ft3gn8vxudum.icu'
- 'ios-hack.icu'
condition: selection
falsepositives:
- Security staff or scanners fetching the hosts on purpose
level: high The better detection is the one the 24 protected organisations already had: violation reports from the browser. A policy in report-only mode costs nothing to deploy and would have reported the first request to an unknown host from real visitors' browsers, on the first page view.
How we test third-party script trust on an engagement
When a web application is in scope, we treat every third-party script as a question about who can change it. We list what the page loads, then ask of each item: which console publishes this, who holds accounts there, how are they authenticated, which API tokens can write, and does a change there pass through anyone's review. The answers are usually worse than the application's own release process, because the tag platform was bought by a different team for a different reason.
Then we test the browser side as an attacker who already owns one of those scripts. On a sanctioned engagement that means standing in for the vendor in a test environment and asking what a hostile file could do on the pages that matter: sign-in, checkout, account settings. Can it read the fields? Can it send data to a host the site has never used? Does anything report it? We are testing the policy and the monitoring, not the vendor.
The findings repeat. A policy that allows eval because a vendor asked for it. A connect-src of a single wildcard. Experiment and tag tools loaded on payment pages where no experiment has run in a year. Publishing rights held by former agency staff. None of these needs an exploit. Each one is the gap that 16 Nov walked through.
What nobody has explained yet
The central fact is missing. Report URI names the candidates as "a compromised Optimizely system, a compromised credential with cross-customer access, or something else", and adds that this is for Optimizely to explain. Hudson Rock says the shared output does not identify whether the access involved provider infrastructure, publishing credentials, integrations or separate customer accounts. Report URI had no response from Optimizely when it published. Until the vendor speaks, every customer should assume its own credentials and tokens are in question and act on that.
The phones are the second open item. Two research teams saw the kit pages being served, and Report URI says the Coruna loader is byte-identical to a sample it analysed in October from a different campaign. Neither team shows a device being exploited. Report URI's advice to anyone who visited an affected site on 16 Nov is to update to iOS 18.7.3 or later, or 26.3 or later, and that is cheap enough to follow without waiting for proof.
The third is the pattern. In October, Report URI documented a campaign it calls DeadTag, which delivered the same two exploit kits through old third-party tags whose expired domains an operator had bought. Different route in, same idea: the script a site forgot it was loading is the delivery network. Three more providers were hit by this operator in the days before Optimizely. A defender should expect the next one to arrive through a vendor that is not on today's list.
What to do today
-
Read what you are serving. Fetch your Optimizely snippet as a visitor would and read the Project JavaScript and every experiment's custom code. Keep a copy with its hash and headers before you change anything.
-
Close the publishing path. Review every user, API token and integration that can write to your projects, remove what you cannot account for, and rotate the rest. The route in is unknown, so assume yours is in question.
-
Treat 16 Nov as exposure where the facts support it. If your own forms take card data on pages that load the snippet, or you keep session tokens in cookies that script can read, follow Report URI's advice: treat them as potentially exposed and rotate the sessions.
-
Fix the policy that would have held. A restrictive script-src without 'unsafe-eval', a connect-src that names only the hosts you use, and reporting turned on. Start in report-only mode this week if you have nothing.
-
Take experiment and tag tools off pages that do not need them. Sign-in, checkout and account pages should load the fewest third-party scripts you can manage.
-
Search your proxy logs for the hosts in the Detection block, back to 13 Nov 2026.
SOURCES
- Report URI // 188 infected snippets across 178 organisations, the timeline in UTC, the Project JavaScript route, the CSP outcomes and the wallet receipts
- Hudson Rock // recovered samples, the embedded payload and its 2.5 second fallback, the other three script providers, and the limits of the iPhone evidence
- Amirul Cyber // independent write-up of the loader, the snippet paths by site and the indicator list
- MDN // connect-src governs fetch, XMLHttpRequest and sendBeacon, the calls the theft code depends on
- Report URI on DeadTag // an October 2026 campaign that delivered the same two iPhone exploit kits through third-party website tags