Subscribe to GEN
Login to GEN
Add a Comment
On 16 November 2026 a compromise at Optimizely put a card skimmer, a crypto wallet clipper and iPhone exploit kits on the websites of 178 organisations, including the BBC, CNN, Zoom and HSBC. How it worked, why so few Content Security Policies stopped it, why third-party scripts are dangerous at the best of times, and why we now block them at the DNS level for our clients.
178 organisations. About 19 hours. A card skimmer, a crypto wallet clipper and iPhone exploit kits, served to visitors of the BBC, CNN, Zoom, HSBC and IKEA from a script every one of them had chosen to trust. Only 24 of the affected sites had a Content Security Policy that stopped it outright.
Not one of those websites' own servers was broken into. They did not need to be.
Optimizely Web is an experimentation service. A website adds one line to its pages, a script tag pointing at cdn.optimizely.com, and marketing staff then use Optimizely's dashboard to test one version of a headline or a button against another. Whatever is saved in that dashboard is delivered to every visitor through the snippet, without anyone touching the website itself.
On 16 November an attacker found a way to save their own code there. Optimizely says an unauthorised party exploited a vulnerability in its Experimentation product, reported to be in the account signup flow, and used it to add JavaScript to customer projects. Optimizely has fixed the flaw, revoked tokens and removed the code.
According to Report URI, which found the attack, the first injection appeared in the BBC's snippet at 04:21 UTC. A card skimmer was added to the payload at 13:43. A mass clean-up began at around 20:40, and by 23:11 every infected file was clean. Report URI checked 1,611 Optimizely snippets and found 188 of them, belonging to 178 organisations, carrying the attacker's code, and expects the real figure to be higher. Optimizely's own login site was among them.
Every one of those files came from Optimizely's own host, with Optimizely's own valid certificate. This was not a lookalike domain or a poisoned cache. The supplier served exactly what it had been given, as it was designed to. Nor was it the attacker's first outing: the same loader turned up in scripts served by Kissmetrics, StorifyMe and Minute Media from 13 November.
This is the part that is easy to miss. A script loaded with a script tag does not run in some sandbox belonging to its supplier. It runs as the page. It can read every field the visitor types into and every cookie the page can read, and it can rewrite anything on the screen, including a wallet address in the middle of a payment instruction.
Adding one hands the keys of your page to the supplier, and to everyone who can ever get into the supplier: their staff, their build systems, their signup flow, and whoever owns their domain after the company moves on. That last one is not hypothetical. In 2024 the polyfill.io domain was sold, and the new owner used it to push malicious redirects through a script that more than 100,000 websites were still loading.
It also changes without telling you. A file on your own server stays as it is until someone deploys a new one. A hosted snippet is whatever its supplier serves this minute, and nothing on your side records the difference.
There is no polite way to put this: mostly, it is convenience. Marketing want A/B testing, analytics, heatmaps, a chat widget, and a tag manager to add more of the same, and every one of them arrives as a snippet to paste. Pasting takes five minutes. Hosting a library yourself, reviewing it and keeping it updated takes an afternoon and a process. So the snippet wins, at organisations with engineering teams perfectly capable of the alternative.
The usual justifications do not stand up. Public CDNs were once said to be faster because a visitor might already have the file cached from another site, but every major browser now keeps a separate cache for each site, so that benefit has gone. A self-hosted copy arrives over the same connection as the rest of the page.
The alternatives are not exotic. A JavaScript library can be downloaded at a fixed version, reviewed, and served from your own domain. Analytics and experimentation both have mature open source tools that run on your own servers. A basic A/B test is a cookie and an if statement on the server, and nobody outside your organisation can rewrite that at four in the morning.
The visitor has never heard of Optimizely and never agreed to trust it. They went to the BBC, or to their bank, and the page they were served was running an attacker's code. The trust they gave was to the name in the address bar, and that is whose reputation pays for it. Speaking for myself, I will not be going back to the BBC website, and I doubt I am alone.
The regulators see it the same way. In 2018 a compromised third-party chat bot on Ticketmaster's payment page took customers' card details, and the Information Commissioner fined Ticketmaster UK £1.25 million, because the page was Ticketmaster's, whoever had written the bot. For anyone taking card payments, PCI DSS version 4 now requires an inventory and authorisation of every script on a payment page and a way of detecting when one changes.
A Content Security Policy is a header a website sends to tell the browser where scripts may come from and where the page may send data. Set properly, it is the browser's backstop against exactly this kind of attack. Most of the affected sites had not set one properly, or at all.
| Content Security Policy on affected sites | What happened | Organisations |
|---|---|---|
| No policy at all | Payload ran and sent data away | 68 |
| Policy too loose to matter | Payload ran and sent data away | 50 |
| Outbound data restricted | Payload ran, could still swap addresses on screen | 17 |
| Payload blocked | Nothing ran | 24 |
Even a careful policy had a problem. Listing cdn.optimizely.com as an allowed source achieved nothing, because the malicious code came from cdn.optimizely.com. Policies using 'strict-dynamic' trusted anything the snippet chose to load. And Optimizely's own documentation tells customers to allow 'unsafe-eval', because some of its features build and run code from text at run time.
The attacker planned for that. One version of the loader carried a complete copy of the payload inside itself, and ran it through eval() if the browser refused to fetch track.js. HSBC's policy blocked track.js and blocked the data leaving, and the inline copy ran anyway.
'unsafe-eval' means what it says. A supplier that needs it is asking for permission to run code it has not shipped yet, which is precisely what an attacker wants as well. If a supplier's documentation tells you to weaken your policy for it, the problem is the supplier.
Subresource Integrity could not have helped either. It pins a script to a known fingerprint, so the browser refuses it if a single byte changes, which only works for a file that is meant to stay the same. A snippet whose whole purpose is to change whenever someone presses save in a dashboard can never be pinned. That is the design, and it is the danger.
Websites are one front. The same pattern, code nobody has read, from people nobody has met, that updates itself, runs right through the tools software is built and hosted with, and it has been a busy autumn.
Each one is the same decision made somewhere else: someone found it easier to trust a third party than to hold the code themselves. We set out our own position in Staging Third Party Modules, and it comes down to one rule. If a package is not trusted enough to be copied onto an internal server and reviewed, it is not trusted enough to run inside the build or production environment. A script tag on a web page deserves no less.
We have said all of this before, and it needs saying again. As of today, cdn.optimizely.com is blocked at the DNS level for all of our clients, along with most of the other third-party script domains that could be used in the same way. A browser that cannot resolve the name cannot fetch the script, whatever the website asks for.
Some websites will lose a feature as a result: an experiment that no longer runs, or a widget that no longer appears. The pages themselves should still load, and a site that cannot show its content without a marketing script has bigger problems. If something you genuinely depend on stops working, raise a ticket and we will look at it.
If you use Optimizely Web, check your Project JavaScript and experiment code for 1688360.cc, atob( and anything you do not recognise, rotate your Optimizely credentials and tokens, and treat card details entered on your site on 16 November as exposed. If you visited any of the affected sites that day, update iPhones and iPads to iOS 18.7.3 or 26.3 or later, check any recent crypto transfers for a changed address, and keep an eye on your card statements.
This article reflects publicly reported information at the time of writing, 18 November 2026, together with our own opinions. Figures come mainly from Report URI, which first found the attack and sells script monitoring, and may be revised as investigations continue. GEN has no commercial affiliation with the vendors named.
This content is not legal or financial advice, and is solely the opinion of the author.