Optimizely Hack Serves Malware to 178 Major Brands

On 16 November 2026, visitors to 178 organisations received malicious JavaScript that had been written into their Optimizely projects. The list of victims reads like a directory of the web: Zoom, the BBC, CNN, Salesforce, GitLab, IKEA, HSBC, Equifax, Air France-KLM, Ledger and Peacock among them. The Optimizely attack did not break into those companies. It broke into the tool they all trusted to run experiments on their pages, and then let that tool serve whatever it was given.
The technical detail matters more than the headline count. This was not a poisoned cache or a hijacked content delivery network. Optimizely's servers delivered exactly the file they were configured to deliver, from Optimizely's own domain, with Optimizely's own certificate. What changed was a field inside each customer's configuration that most people never look at: the slot where a marketing team can paste custom JavaScript. Whoever gained access to that slot across dozens of unrelated accounts turned a trusted include into a delivery mechanism for a multi-purpose malware kit. The result combined cryptocurrency address swapping, card skimming, cookie and screenshot theft, and two iOS exploit kits in a single script. Cryptocurrency had already been stolen and moved before the clean-up finished. This article sets out what the Optimizely attack actually delivered, how it worked, who was caught, how much money is accounted for, and what the response looked like hour by hour. It also explains why the usual advice about trusting well-known vendors failed here, and what site owners can do that would have helped.
What the script actually did
Every visitor received the same file, a script called track.js served from an attacker domain. Every part of it was active at once. What a person experienced depended on their device and on what they did while the page was open.
The first function was cryptocurrency address swapping. The script searched for a valid wallet address across Bitcoin, Ethereum and EVM chains, TRON, Solana, TON, Litecoin, Bitcoin Cash and Dogecoin, checked the checksum so that only genuine addresses were replaced, and substituted one controlled by the attacker. It did this in page text, in form fields, in links and in images. It also did it in the clipboard, both when a visitor copied and through direct clipboard access, and in outgoing requests, new windows, navigation events and the iOS share sheet. It re-scanned the page whenever the content changed, so a page that loaded clean could become hostile seconds later.
The second function was card skimming, added partway through the day. The script watched form submissions and every outbound POST request for a card number, checked with the Luhn algorithm, together with an expiry date and a security code. When it had all three it sent them to the attacker's server. Cards typed into a hosted payment iframe from a provider such as Stripe or Adyen were out of reach. Cards typed into a site's own form fields were not.
The third function was theft of session material. When the script swapped an address next to wallet or payment wording, it also sent back the page address, the contents of the cookie store for anything not marked HttpOnly, the original addresses it had replaced, and a rendered screenshot of the page. That combination gives an attacker a picture of what the victim saw alongside the identifiers needed to act as them.
The fourth function targeted phones. On iPhones and iPads the script read the operating system version and loaded an exploit chain in a hidden frame, matching the chain to the device. Older versions received Coruna, the kit that Google's threat intelligence group documented in March 2026. Versions in the 18 range received a separate chain referred to as DarkSword. A third chain aimed at version 26 appeared to be fingerprinting only. Pages visited on iOS 18.6 or earlier reloaded every 60 seconds to retry the exploit, which is a visible symptom someone could have noticed on an affected phone.
How the Optimizely attack worked
Optimizely Web is an experimentation and personalisation product. Teams use it to run A/B tests, target content at audience segments, and switch features on or off without a deployment. Because it has to alter the page for each visitor, it has to run code in the browser, which means every participating site carries an Optimizely script. That script is the product. Nothing unusual was required for it to reach a visitor.
The platform runs experiments by having each customer load a JavaScript snippet from Optimizely's content delivery network. The file contains Optimizely's own code, the customer's configured experiments, and an optional field called Project JavaScript. That field exists so a customer can add custom behaviour that runs on every page, and it is edited through the Optimizely dashboard.
The attacker wrote code into the Project JavaScript of customer projects. The rest of each snippet stayed untouched. The file was still served from the vendor's network, with the vendor's certificate, from a host that every affected site had already chosen to trust. One Optimizely snippet examined during the response contained a small block that created a script element, pointed it at an attacker domain, and appended it to the page.
That is the whole mechanism, and it is why the Optimizely attack is best described as a supply chain problem rather than a website compromise. The company whose site was affected did nothing wrong with its own code. It had delegated a piece of its page to a third party, and the third party's configuration surface had been altered.
Some customers do not load the snippet from Optimizely's network at all. They mirror it under their own domain, which looks safer because the file appears to be first party. It was not safer. Those sites were affected too, because the content still came from the same Optimizely project. Mirroring changes where a file is served from. It does not change who controls what the file contains.
The mechanism inside Optimizely is now partly described by the vendor itself. In a notice posted to the Optimizely service status page at 01:02 UTC on 17 November under the title Web Experimentation Security Incident, the company said an unauthorized party had exploited a vulnerability in Experimentation and used it to add malicious JavaScript to some customer projects. It said the vulnerability was fixed, the injected code removed from affected projects, affected user accounts secured, and additional protections put in place to prevent unauthorized code from being delivered.
The Optimizely statement confirms the shape of the failure without explaining it. It does not describe how the vulnerability worked, how many projects were modified, when the flaw was introduced or fixed, whether the attacker retains any access, or what customers should do next. The outside analysis could not determine the access path either, and explicitly ruled out cache poisoning. Optimizely's own domains carried the code as well, which argues against the idea that a single customer account was the entry point.
Three loaders and a moving target
The code injected into the Optimizely snippets was not static. The response identified three variants of the loader, and the differences show an operator adapting to defences in real time.
The simplest variant was a plain script tag pointing at the attacker domain. That is easy to detect with a search for the domain string. The second variant hid the same address in base64 encoding and decoded it at runtime, so a text search for the domain found nothing. One company's snippet was cleaned and then re-infected with this variant, which indicates the attacker still held access at that moment. The third variant embedded a full copy of the payload inside the snippet and ran it through the language's evaluation function if the external domain was blocked.
The payload itself was rebuilt 26 times during the day. Features were added and wallet addresses were added, but nothing was removed. The card skimmer joined at 13:43 UTC, roughly nine hours after the first injection. That pattern of incremental accumulation is characteristic of an operator watching results and expanding scope while the window is open.
For defenders, the lesson is that string matching against a known bad domain is a temporary control. A base64 encoded address, an embedded payload, or a newly registered domain defeats it. Detecting unexpected change in a trusted file, and constraining what that file is permitted to do, survive those variations.
The iPhone branch
The mobile component deserves separate attention because it changed the risk profile of the incident. A desktop visitor faced address swapping and, later, card theft. A visitor on an outdated iPhone faced an attempt to take over the device.
The exploit chains were not written for this campaign. The Coruna loader observed here was byte-identical to a sample analysed in October from a different operation, and Google's earlier research traced the kit through a surveillance vendor customer, watering hole attacks against Ukrainian users, and later mass campaigns run by a financially motivated group. The kit's own comments state that iOS 18.7.3 and later, and 26.3 and later, are patched.
That detail is the reassuring part and the uncomfortable part at once. Apple had already fixed the underlying vulnerabilities. Anyone running current software was not exposed to the mobile chain. The exposure fell entirely on people who had not updated, sometimes because the hardware could no longer take a current release. The Optimizely attack therefore converted a routine software maintenance gap into a device takeover risk for a slice of visitors who had no way to know what they were loading.
Who was caught
The response fetched 1,611 distinct Optimizely snippets and checked each for the malicious code. Of those, 188 snippets belonging to 178 organisations were infected. A copy of every infected file was retained along with response headers, a hash and a fetch timestamp as evidence. The published figure is explicitly a floor, because only snippets that could be located in time were checked.
The sector breakdown shows how deeply experimentation tooling reaches. Banking and finance accounted for 32 organisations, including HSBC entities in the United States, Hong Kong, Mexico, Singapore, Australia, the United Arab Emirates, Malaysia, Taiwan, Vietnam, Egypt and the Philippines, which is 11 markets for one brand. Equifax appeared through several domains, alongside ABN AMRO, S-Pankki, ICS Cards, Rest Super, Hang Seng Investment and TaxAct. Technology and software accounted for 23, with Zoom, Bitly, Salesforce, GitLab, Mixpanel, Typeform, Tableau, Hootsuite, Smartsheet, MuleSoft and Optimizely itself on the list.
Retail and fashion added 20, including IKEA, ASOS, Mattel, American Eagle and Levi's across a spread of national storefronts. Streaming and entertainment contributed 12 with Peacock, NOW TV, Zattoo, Starz and SkyShowtime. News and media contributed 10, among them CNN, the BBC, Fox News, NPR, CTV News, BNN Bloomberg and Top Gear. Airlines and travel added nine, including Alaska and Hawaiian Airlines, Transavia, Air France-KLM and KLM in the Netherlands.
The remaining sectors covered healthcare, industrial manufacturing, telecoms, cryptocurrency and trading, insurance, property, food and hospitality, home and smart home, nonprofit, gambling, professional services, logistics, automotive and education. Twenty sectors in total. For a reader trying to judge personal exposure, the practical signal is not the brand list but the categories: if a site ran experiments through this platform and accepted payment details or hosted a wallet workflow, it was a candidate.
The money so far
As of 19:25 UTC on 16 November, the attacker's wallets had received about 7,040 dollars, all within roughly nine hours. The largest holding was 0.08298 Bitcoin across four transactions, valued at about 5,380 dollars and still unspent at the time of publication. A Litecoin address held 17.02 coins worth about 762 dollars. A TRON address held 662.9 USDT and 184 TRX, about 719 dollars. An Ethereum address held 0.055 ETH and the same address on BNB Chain held 82 USDT.
Funds were moved out between 13:03 and 13:05 UTC. Stablecoin on Ethereum went to a deposit address at a centralised exchange. The TRON wallet had been in use since March 2026, which places this operator in the ecosystem well before this campaign.
Two caveats belong next to those figures. The visible total covers cryptocurrency only, and the response could not see how many card details or cookie sets reached the collection endpoint. The true financial impact is therefore unknown and probably larger. And the sum is small relative to the brand list, which suggests either that the campaign was still ramping up when it was cut off or that the operator's economics depend on volume across many victims rather than a single large haul.
Twenty-six rebuilds in one day
The timeline is unusually complete and worth reading as a case study in how fast these operations move.
On 16 November the last clean capture of the BBC's snippet was at 02:21 UTC. The first injection appeared at 04:21, when the BBC snippet began loading an earlier attacker domain. Certificates for the second domain were issued between 04:28 and 05:23, and the domain itself was registered at 05:23. The first build of the crypto swapping script appeared at 08:10 and was observed on the BBC at 08:28. Zoom's snippet was modified at 11:20. Stolen cryptocurrency was swept to exchanges at 13:03. The card skimmer was added at 13:43. A second iOS routing domain was registered at 16:38, and between 17:13 and 17:57 the mobile exploit frame was injected into visits from iOS 18.6 and earlier.
The discovery sweep ran between 19:24 and 19:59 and found 186 of 1,611 snippets infected. Mass clean-up followed between roughly 20:40 and 21:10, with Ledger cleared at 20:41, Smartsheet at 20:56 and the BBC at 21:07. By 21:52, one of 188 snippets remained infected, belonging to TalkTalk.
The earlier context matters too. The same injector domain had appeared on Turkish sites from 17 September through what was described as an SEO backlink panel script. Between 13 and 16 November a prior wave used a different injection path through three separate marketing and media scripts, hitting sites including securitybank.com, faz.net, smh.com.au, vidaxl and foxsports.com. The Optimizely attack was not the opening move. It was the escalation to a platform with far greater reach.
The sequence had a second act. At 22:08 UTC Optimizely said it was continuing to investigate. By 23:11 all infected Optimizely files were reported clean. The collection domain was placed on clientHold at 00:37 UTC on 17 November and stopped resolving, and the earlier injector domain followed at 03:37. Between 04:11 and 05:39 the attacker's wallets were emptied into new addresses.
Estimates of the exposure window differ because they measure from different points. One independent analysis puts it at at least 13 hours, counting from the first payload fetch to the moment files were cleaned, while a consolidated timeline puts it at nearly 19 hours, counting from the first injection. Both are defensible readings of the same evidence, and neither should be presented as the settled figure.
Why the network was not the problem
It is tempting to read this as a content delivery network failure, because the malicious file arrived from an Optimizely content delivery network. That reading is wrong on the specifics and unhelpful on the lesson.
The network served what it was given. The file was built from customer configuration, and the configuration had been altered. No cache was poisoned, no certificate was forged, no routing was hijacked. Every trust signal a browser checks was valid, because the file genuinely came from the expected place.
That is the uncomfortable structural point. The web runs on delegated code. A typical publisher loads scripts for analytics, advertising, consent, personalisation, customer support and experimentation, each from a vendor it does not control. Each of those vendors has an administrative surface, and each surface is a way in. The Optimizely attack is a demonstration that compromising the administrative layer of a widely used platform produces malware on 178 sites without touching any of them. The platform does not even need to be the weakest link in a publisher's stack. It only needs to be the one with the most reach.
For related reading on how third-party trust fails in practice, our account of the Rank Math PRO backdoored update covers a similar shape of incident in a different part of the stack, and our coverage of the SAP NetWeaver critical flaws looks at what happens when the underlying platform is the weak point.
What would have helped
The published Optimizely analysis is specific about controls, and they fall into two groups: things that would have blocked parts of this campaign and things that would have revealed it.
On blocking, the published data is more useful than the generic advice, because the response tested the content security policies of 256 affected hosts against what the payload needed. Sixty-eight had no policy at all. Fifty had a policy that permitted both the payload and the data theft. Seventeen ran the payload but blocked the outbound connection, which stopped card, cookie and screenshot theft while address swapping still worked. Twenty-four, about 15 percent, blocked the payload outright.
Two of those findings cut against common assumptions. Allowlisting the vendor's content delivery host achieved nothing, because the malicious code arrived from that host. What separated the protected group was a script policy listing only the hosts a page genuinely needed. And both the strict-dynamic directive and permission for runtime evaluation reopened the hole: the first extends trust to any script the trusted snippet inserts, and the second let the embedded fallback run. The analysis notes that the vendor's own documentation tells customers to permit runtime evaluation. At one bank, the policy blocked the external script and the exfiltration but the inline copy still executed.
Subresource integrity is not a practical control for a file that legitimately changes per customer. Integrity monitoring of the served file, by contrast, is exactly the signal that was missing.
On detection, the useful signals are change in a trusted file, a new script appearing on a page, and a connection to an unfamiliar destination. A snapshot of each third-party script at a known good moment, compared against what is served later, would have surfaced this quickly. So would monitoring for new external hosts contacted by the browser. Both are ordinary engineering work, and both were the difference between finding this in hours and finding it in weeks. In this case the detection and the clean-up came from outside Optimizely, which is worth noting on its own.
There is also a governance point that no scanner fixes. The Project JavaScript field is powerful, which means access to it is equivalent to the ability to run code on the site. That access should be treated like production deployment rights: limited to named people, reviewed, logged and revocable. If a marketing team can paste arbitrary script into a production page, the marketing login is a production credential. For teams that want continuous visibility of what their pages actually load, a monitoring product such as the one we reviewed in our look at AI-powered website security monitoring covers the same ground from the defensive side.
There is a further gap that no control addresses. No affected company has issued a public statement about the incident, and no national cyber security centre, payment card industry body or card network had published a notice at the time of writing. Customers of the banks, retailers and airlines on the list therefore learned about their exposure from a third-party security blog rather than from the organisations that hold their data.
Questions readers are asking
Was Optimizely itself breached? Optimizely states that an unauthorized party exploited a vulnerability in its Experimentation product, which gave access to customer project settings. It has not described the vulnerability, and the independent analysis could not determine the access path. What is established is that the content delivery network served files exactly as configured, so this was not a cache poisoning event.
Did the Optimizely attack steal payment cards? A card skimming module was added at 13:43 UTC and was built to capture card numbers, expiry dates and security codes entered into a site's own form fields, sending them to the attacker. Cards entered inside hosted payment frames from providers such as Stripe or Adyen were out of its reach. How many card details were actually collected is not known, and no card theft has been confirmed publicly.
Are iPhone users at risk? Only those who had not installed the relevant updates. The mobile chains relied on vulnerabilities that had already been patched, and the kit's own notes state that iOS 18.7.3 and later, and 26.3 and later, are fixed. Current devices were not exposed to that part of the campaign.
How much money was taken? About 7,040 dollars in cryptocurrency was visible in the attacker's wallets as of 19:25 UTC on 16 November. The volume of card and cookie data collected is not known from public information, so the total loss is larger than the visible figure.
Which companies were affected? 178 organisations across 188 snippets, spanning 20 sectors. Named examples include Zoom, the BBC, CNN, Salesforce, GitLab, IKEA, HSBC, Equifax, Air France-KLM, Ledger and Peacock. The published list is a floor rather than a complete inventory.
What should a visitor do? Anyone who entered card details on an affected site during the window should contact their card issuer. Anyone who copied a wallet address from an affected page should verify it against a known good source before sending funds. Anyone on an older iPhone who visited an affected site should update the device, and if a wallet was installed, move funds to a new wallet with a fresh recovery phrase generated on a clean device.
Sources
- Report URI, World's Largest Brands Infected With Malware Via Optimizely
- Optimizely service status dashboard
- Archived capture of the Optimizely incident notice
- StepSecurity, independent analysis of the injected scripts
- SOCRadar, consolidated incident timeline
- Google Threat Intelligence Group, Coruna iOS exploit kit research
- Optimizely trust center, security