Back to articles

Optimizely Web snippets served a card skimmer and crypto clipper on 178 brands' sites

What happened

On November 16, 2026, malicious JavaScript reached visitors of at least 178 organizations through Optimizely Web, the experimentation and A/B testing platform that many large sites load on every page. Security researcher Scott Helme's team at Report URI found the code in 188 of the 1,611 Optimizely snippets it checked, on sites including Zoom, the BBC, CNN, Salesforce, GitLab, IKEA, HSBC, Equifax, Air France-KLM, Ledger and Peacock (Report URI). StepSecurity independently confirmed the loading chain on a GitLab page and a Mutual of Omaha page, and PrivacyAffairs found the same injected line in six of the ten Optimizely files it downloaded that evening, including those used by CNN, O2 and Optimizely's own website (StepSecurity, PrivacyAffairs).

The payload did several things at once. It swapped cryptocurrency wallet addresses on the page and in the clipboard for the attacker's, captured card numbers typed into a site's own payment forms, sent cookies and page screenshots to the attacker and, on iPhones and iPads, loaded exploit kits known as Coruna and DarkSword (Report URI). Report URI counted about $7,040 in the attacker's wallets by 19:25 UTC, and StepSecurity put the figure at about $6,600. Neither could see how many cards, cookies or phones were affected.

Where Outtake fits. The injected line ran inside trusted pages, served from Optimizely's own servers, and catching that change in the browser is the job of content security policy and script monitoring. Everything the line called, though, lived on the open internet: a skimming and address-swapping server registered that morning, iOS exploit routers on throwaway domains and an injector domain used in earlier waves. That shared attacker infrastructure is what Outtake's agents are built to find, trace to related domains and report to the registrars and hosts that can take it offline. In this incident it was the one thing all 178 sites had in common, and registry records we checked show the main server's domain was not placed on hold until about 19 hours after it was registered. The rest of this article covers what happened, then where outside monitoring could have helped and where it could not.

How the code got into trusted snippets

Optimizely Web works by having each customer load a JavaScript file, called a snippet, from cdn.optimizely.com. The snippet holds Optimizely's code, the customer's experiments and an optional field called Project JavaScript, which is custom code the customer writes in the Optimizely dashboard and which runs on every page. According to Report URI, the attacker wrote a short loader into the Project JavaScript of customer projects and left the rest of each snippet untouched. The file still came from Optimizely's domain, with Optimizely's certificate, from a host every affected site already trusted. Customers who serve their snippet from their own domain were affected too, because the content comes from the same project.

Report URI saw three versions of the loader. Most snippets carried a plain script tag pointing at the attacker's server. Some hid the same address in Base64 so a simple search for the domain would miss it. A third version carried a full inline copy of the payload that runs through eval if the server is blocked. Smartsheet's snippet was cleaned at one point and then re-infected with the hidden version, which Report URI reads as a sign that whoever wrote the code still had access at that time.

How the attacker got write access to so many unrelated projects is not known. Report URI says the possibilities include a compromised Optimizely system or a credential with cross-customer access, and that this was not a CDN cache-poisoning attack: the CDN served exactly what the projects contained. Optimizely's own snippets for www.optimizely.com and login.optimizely.com carried the code as well. As of late on November 16, Optimizely had published no statement, according to Report URI, StepSecurity and MATT Research (MATT).

What the payload did to visitors

Every visitor to an infected page received the same script, and what happened next depended on the device and on what the visitor did.

  • Crypto address swapping. The script looked for valid wallet addresses for Bitcoin, Ethereum and other EVM chains, Tron, Solana, TON, Litecoin, Bitcoin Cash and Dogecoin, and replaced them with the attacker's in page text, form fields, links, outgoing requests and the clipboard. PrivacyAffairs reports it rescanned the page every 50 milliseconds. Copying an address while the tab was open could paste the attacker's instead.
  • Card skimming. Report URI says this feature was added at 13:43 UTC. The script watched form submissions and outgoing requests for a card number that passes the standard checksum, an expiry date and a security code, then sent all three to the attacker. Card fields hosted inside a payment provider's own frame, such as Stripe or Adyen, were out of its reach. Card fields built into a site's own forms were not.
  • Cookie and screenshot theft. When the script swapped an address near wallet or payment wording, it sent the page address, any cookies not marked HttpOnly, the original addresses and a screenshot of the page.
  • iOS exploit kits. On iPhones and iPads the script checked the iOS version and loaded an exploit chain in a hidden frame from separate .icu domains. Report URI notes that the kit's own comments describe iOS 18.7.3 and later, and 26.3 and later, as patched, and that pages on iOS 18.6 or earlier reloaded every 60 seconds to retry, which is a visible symptom.

The infrastructure behind it

Report URI's timeline, combined with registry records we retrieved at about 01:27 UTC on November 17, shows an operator that staged infrastructure in advance and kept building through the day. Domain names here are written in defanged form and are not linked.

  • webshell[.]lol, which Report URI ties to the same actor's injector, first appeared on Turkish sites through an SEO "backlink panel" script on September 17. Between November 13 and 16, an earlier wave injected it through Kissmetrics, StorifyMe and Minute Media scripts. It is also what the BBC's snippet loaded in the first Optimizely injection Report URI saw, at 04:21 UTC on November 16. The registry record shows it was registered in April 2024, sits behind Cloudflare and still resolved at our check (registry record).
  • 5vobauua3ojpklx[.]icu, an iOS exploit router, was registered on November 9 at 17:52 UTC, a week before the Optimizely wave (registry record). StepSecurity reports a related iOS-stage domain was serving a page by November 12.
  • 1688360[.]cc, the server that delivered the payload and received stolen cards, cookies and screenshots, was registered through Spaceship at 05:23 UTC on November 16. Report URI first saw its script on bbc.com at 08:28 UTC. The record we retrieved shows a client hold, last changed at 00:37 UTC on November 17, and the name no longer resolved (registry record). The record does not say who requested the hold.
  • eil786npto9u5bq[.]icu, a second iOS router, was registered at 16:38 UTC on November 16 while the attack was live (registry record). Both .icu routers use the same registrar name servers, still resolved at our check and showed no hold. PrivacyAffairs lists a third .icu host among the domains to block.

Report URI says the payload was rebuilt 26 times during the day, and that the attacker swept stolen crypto to exchanges at about 13:03 UTC. Its sweep between 19:24 and 19:59 UTC found 186 infected snippets out of 1,611. After a mass clean-up between about 20:40 and 21:10 UTC, one infected snippet, TalkTalk's, remained at 21:52 UTC. StepSecurity estimates an exposure window of at least 13 hours.

Who was affected

Report URI's list of 178 organizations spans most industries. Banking and finance account for 32, including several HSBC regional sites, Equifax and Paychex. Technology and SaaS account for 23, including Zoom, Salesforce, GitLab and Optimizely itself. Retail and fashion account for 20, followed by streaming, news and media, airlines, healthcare, telecoms, crypto and trading, insurance and others. Report URI says the real number is likely higher, because it could only check snippets it found in time, and PrivacyAffairs makes the same point about its smaller sample.

Report URI also checked how much content security policy helped, by crawling 256 of the affected organizations' own hosts. In its results, 68 organizations had no policy at all and 50 had a policy that allowed both the payload and the data theft. For 17, the payload ran but a restrictive connect-src setting blocked card, cookie and screenshot theft, though address swapping still worked inside the page. Twenty-four blocked the payload. Allowing cdn.optimizely.com did not help, because the code arrived from that host. At HSBC, the policy blocked the external script and the data theft, but the inline copy still ran because the policy allowed eval, which Report URI says Optimizely's documentation tells customers to permit.

What remains unknown

  • How the attacker gained write access to customer projects, and whether that access has been fully removed.
  • How many cards, cookies and screenshots reached the attacker, and how many iOS devices were exploited.
  • Who is behind the campaign. Report URI notes one of the attacker's wallets has been in use since March 2026, so the actor was active before this wave.
  • The exact count of affected sites. Report URI reports 188 snippets across 178 organizations, StepSecurity cites 186 sites and the two put the stolen crypto at different totals at different times.
  • Who requested the registry hold on the main server's domain.

Why this matters beyond Optimizely

The same actor used at least four different third-party script providers in a week, according to Report URI. Any vendor that can change code running on a brand's pages becomes a way to reach that brand's customers, and none of the affected organizations had to be breached themselves.

The harm still lands on the brand. A customer whose card is misused, or whose crypto went to the wrong wallet, saw the brand's own domain in the address bar. For security, fraud and communications teams, the useful question is what they could see and act on from outside while the code was live, and how quickly the infrastructure behind it could be cut off.

Where Outtake could have helped

This section is hypothetical analysis, kept separate from the reported facts above. It does not suggest that any affected organization uses or lacks Outtake. Outtake does not inspect the contents of a site's own third-party scripts and could not have stopped the attacker from writing into Optimizely projects. It maps the externally visible parts of this attack to capabilities Outtake describes publicly.

Tracing one indicator to the whole setup

Outtake's Platform page describes a Recon Agent that takes a signal back to its origin and expands it across the wider adversarial network, and its Brand Protection page says agents trace shared hosting, registration patterns and linked accounts from a single flagged domain. Public scans recorded requests to 1688360[.]cc on major sites from 08:28 UTC. Starting from that one domain, an investigation could support finding the rest of the setup: the .icu routers registered on November 9 and 16, the injector domain from the earlier wave, and the other script providers the same actor used.

One caution applies. The two .icu routers share registrar name servers, and shared defaults like that are not proof of a single operator on their own. Tracing would need to lean on the payload's own references and on timing as well.

Getting registrars and hosts to act

Outtake describes evidence packaging, automated reporting and platform escalation, with remediation through registrars and hosts. In this case the registry record for 1688360[.]cc shows a hold about 19 hours after registration and about three and a half hours after the mass clean-up of snippets. A hold on that domain cuts off the plain and Base64 loaders on any snippet not yet cleaned, which mattered here: TalkTalk's snippet still carried the code at 21:52 UTC. A report that ties the domain to the payload, the stolen-data endpoints and public scan evidence gives an abuse desk what it needs to decide. Our guides to taking down a phishing or lookalike domain and dismantling the operation behind fake domains walk through those steps.

The iOS routers are a separate case. At our check they still resolved with no hold, and Report URI ties them to an exploit chain that can reach users outside this incident.

Watching for the stolen data and the aftermath

Outtake's platform page lists the dark web among the surfaces it covers. Stolen card details and session cookies are the kind of data that can later be offered for sale, and monitoring for a brand's domains and customer data in those markets could support a faster fraud response. Our explainer on dark web brand monitoring covers what that monitoring finds. No public source we found says data from this incident has been offered for sale.

Public incidents can also invite scams that borrow the brand and the news, such as fake security notices, refund offers or crypto recovery services. That has not been reported in this case. Outtake's Brand Protection monitoring of lookalike domains, impersonation accounts and fraudulent ads is the relevant capability if it appears.

Getting the indicators to the people who block them

Outtake says its threat intelligence flows into SIEM, SOAR and ticketing tools. Indicators like the attacker's domains are what a security team needs to add to content security policy, proxy and DNS blocklists. Report URI's data shows that a restrictive policy blocked the external payload at 24 organizations. Getting the right domains into those controls quickly is part of the response, even for teams whose own snippet was never infected.

Trace the Optimizely injection from snippet to registry hold

Choose a path, then step through its stages. Each stage shows what sources reported, where external monitoring could have intervened (hypothetical analysis, not a reported event) and the boundary of that intervention. Times are UTC and come from Report URI, StepSecurity, PrivacyAffairs and registry records we retrieved at about 01:27 UTC on November 17, 2026.

What a visitor's browser did

Snippet

A loader is written into the site's Optimizely Project JavaScript

Reported evidence

The attacker added a short loader to the Project JavaScript of customer projects; the rest of each snippet was untouched and still served from cdn.optimizely.com (Report URI).

Possible external intervention

Little outside intervention is possible here. The change happens inside a vendor account and a trusted file, which is where script monitoring and content security policy work.

Control boundary

Outtake does not inspect a site's own third-party scripts or vendor accounts. Access reviews and credential rotation are the site owner's controls.

Loader

The page fetches the payload from 1688360[.]cc

Reported evidence

Most snippets loaded a script from the attacker's server; some hid the address in Base64, and one version carried an inline copy that runs through eval if the server is blocked (Report URI).

Possible external intervention

Hypothetical: the attacker's domain is visible on the open internet and in public scan records, so it can be investigated and reported to its registrar and host.

Control boundary

Taking the domain offline does not stop the inline eval version, which needs no external server.

Clipper

Wallet addresses on the page and clipboard are swapped

Reported evidence

The script replaced valid wallet addresses for many chains in page text, fields, links, requests and the clipboard (Report URI, PrivacyAffairs). Funds were swept to exchanges at about 13:03 UTC.

Possible external intervention

Hypothetical: sharing the attacker's wallets and domains with fraud and support teams could help them recognize customer reports quickly.

Control boundary

Swapping happens inside the visitor's page. Recovering on-chain funds depends on exchanges and law enforcement, not on monitoring.

Skimmer

Card numbers typed into site forms are sent to card.php

Reported evidence

Added at 13:43 UTC. The script captured card number, expiry and security code from a site's own forms; fields inside hosted payment frames were out of reach (Report URI).

Possible external intervention

Hypothetical: dark web monitoring for the brand's domains and customer data could support a faster fraud response if stolen cards are later offered for sale.

Control boundary

No public source says data from this incident has been offered for sale, and how many cards were taken is unknown.

iOS kit

iPhones and iPads load an exploit chain from .icu domains

Reported evidence

The script checked the iOS version and loaded Coruna or DarkSword chains in a hidden frame; vulnerable pages reloaded every 60 seconds to retry (Report URI).

Possible external intervention

Hypothetical: the .icu routers are separate domains that can be traced and reported, which matters because they can serve users outside this incident.

Control boundary

Device patching is the user's and the device maker's control. Monitoring cannot see whether a given phone was exploited.

What the infrastructure shows

Sep 17 to Nov 16

An injector domain appears in earlier waves

Reported evidence

Report URI ties webshell[.]lol to the same actor: first seen September 17 via an SEO backlink script, then injected through Kissmetrics, StorifyMe and Minute Media scripts from November 13. It was registered in April 2024 and still resolved at our check (registry record).

Possible external intervention

Hypothetical: tracing this domain across earlier waves could have connected it to the operator before the Optimizely wave.

Control boundary

Report URI's link between the waves is its own analysis. We did not visit the domain.

Nov 9

The first iOS router is registered a week early

Reported evidence

5vobauua3ojpklx[.]icu was registered on November 9 at 17:52 UTC (registry record). StepSecurity reports an iOS-stage domain was serving a page by November 12.

Possible external intervention

Hypothetical: pre-staged infrastructure is what network tracing aims to expose before it is used.

Control boundary

A random-looking name carries no brand signal, so it is found by tracing from another indicator, not by brand monitoring alone.

Nov 16, 05:23

The payload server is registered hours before use

Reported evidence

1688360[.]cc was registered through Spaceship at 05:23 UTC; its script was first seen on bbc.com at 08:28 UTC (registry record, Report URI).

Possible external intervention

Hypothetical: from the first public scan, evidence tying the domain to the payload and its theft endpoints could be packaged for the registrar and host.

Control boundary

Removal is the registrar's and host's decision, and the inline loader variant would keep working without this domain.

Nov 16, 16:38

A second iOS router is added while the attack is live

Reported evidence

eil786npto9u5bq[.]icu was registered at 16:38 UTC on November 16 and added to the iOS stage that evening (registry record, Report URI).

Possible external intervention

Hypothetical: a new domain referenced by a known payload can be picked up by watching the payload, not only the registration feed.

Control boundary

Shared registrar name servers with the first router are a default, not proof of one operator on their own.

Nov 16, 20:40

Snippets are cleaned, one remains

Reported evidence

A mass clean-up ran from about 20:40 to 21:10 UTC; at 21:52 UTC one of 188 infected snippets, TalkTalk's, still carried the code (Report URI).

Possible external intervention

Hypothetical: while any snippet remains infected, cutting off the payload server protects that site's visitors too.

Control boundary

Clean-up of the snippets themselves is up to Optimizely and each customer.

Nov 17, 00:37

The payload domain goes on hold; the iOS routers do not

Reported evidence

At our check, 1688360[.]cc showed a client hold last changed at 00:37 UTC and no longer resolved. Both .icu routers still resolved with no hold (registry record).

Possible external intervention

Hypothetical: continued reporting of the remaining routers with the exploit evidence could support further holds.

Control boundary

The record does not say who requested the hold, and we did not test whether the routers still serve the exploit.

What outside monitoring cannot reach

Most of this attack happened where external monitoring has no view. The write into Optimizely projects took place inside vendor accounts, and the payload ran inside visitors' browsers on the brand's own pages. Detecting that change is the job of script monitoring and content security policy, and preventing it depends on who holds write access to the vendor account. Those are the site owner's controls.

Taking down infrastructure has its own limits. The inline version of the loader carried the whole payload and needed no external server, so a registry hold would not have neutralized it. Registrars, registries and hosts decide whether to act, and the record for the payload server does not say who requested its hold. Stolen crypto had already been moved to exchanges by early afternoon, where recovery depends on the exchanges and law enforcement.

Several facts remain unknown: how many people were affected, what data the attacker received and who runs the operation. This analysis relies on the researchers' published findings and on our own registry and DNS lookups. We did not visit the attacker's domains or run the payload.

Frequently asked questions

Was Optimizely hacked?

That is not established. Report URI says the attacker wrote code into customer projects and that the CDN served exactly what the projects contained, so this was not CDN cache poisoning. How the attacker got write access, whether through an Optimizely system or a credential with cross-customer reach, is unknown, and Optimizely had published no statement as of late on November 16 (Report URI, MATT).

What should a company that uses Optimizely Web do now?

Report URI recommends checking every project's Project JavaScript and experiment code for the attacker's domain, atob( and any script you did not write; reviewing and rotating users, API tokens and integrations with write access; treating cards entered in your own forms on November 16 as potentially exposed; rotating secrets in cookies not marked HttpOnly; and tightening content security policy, especially connect-src and unsafe-eval (Report URI). StepSecurity notes a clean file today does not prove it was clean earlier (StepSecurity).

Why didn't allowlisting Optimizely in content security policy help?

The malicious loader arrived inside the trusted Optimizely file itself. Report URI found that what protected the 24 sites that blocked the payload was a restrictive script-src that did not allow the attacker's server, while strict-dynamic and unsafe-eval let the attack through. A restrictive connect-src blocked data theft at 17 more sites, though address swapping still worked (Report URI).

Could Outtake have stopped this attack?

Not the injection itself. Outtake does not inspect the contents of a site's own third-party scripts or vendor accounts. What it addresses is the external part: tracing the attacker's domains, packaging evidence for registrars and hosts, and watching for stolen data and brand abuse afterward. A hold on the payload server cuts off the external loaders, but not the inline version that runs through eval.

Is the attacker's infrastructure offline?

Partly. At about 01:27 UTC on November 17, the registry record for the payload server's domain showed a client hold and the name no longer resolved, while both .icu iOS routers and the injector domain still resolved with no hold (1688360[.]cc record, 5vobauua3ojpklx[.]icu record, eil786npto9u5bq[.]icu record). Resolving is not proof a domain still serves the exploit; we did not visit them.

Sources