Optimizely Web Experimentation Supply-Chain Attack: Prepare Now

Optimizely Web Experimentation security advisory: vendor script hijacked, prepare now. High, no CVE. Not in CISA KEV.

Status: Optimizely says an unauthorized party exploited a vulnerability in Web Experimentation to add malicious JavaScript to some customer projects. It says it has fixed the vulnerability, removed the injected code, secured affected user accounts and notified affected customers (status page and statement to BleepingComputer). Its status page entry, posted at 20:02 EST on 16 November 2026, was still marked Monitoring when we checked at 04:41 EST on 18 November. Report URI says every infected snippet it tracked was clean by 18:11 EST on 16 November. On 17 November Optimizely issued an interim incident report to customers, with a preliminary timeline, the attack vector and preliminary indicators; it says a final root-cause analysis will follow Mandiant's investigation. No CVE has been published, and Optimizely has not said how many projects were changed.

Executive Summary

Optimizely Web Experimentation runs A/B tests through a JavaScript "snippet" each customer site loads from Optimizely's CDN. It carries the customer's experiments and an optional Project JavaScript field: custom code written in the Optimizely dashboard that runs on every page.

According to Report URI's analysis, written by researcher Scott Helme:

  • An attacker wrote a loader into customer projects' Project JavaScript; the files were still served from cdn.optimizely.com with Optimizely's certificate.
  • Report URI checked 1,611 Optimizely snippets and found the code in 188, belonging to 178 organisations. It says the real number is likely higher, because it could only check the snippets it found in time.
  • Named organisations include Zoom, the BBC, CNN, Salesforce, GitLab, IKEA, HSBC, Equifax, Air France-KLM, Ledger and Peacock. Optimizely's own www.optimizely.com and login.optimizely.com snippets carried the code too.
  • The second-stage script swapped cryptocurrency wallet addresses, skimmed card data typed into a site's own form fields, sent cookies and screenshots when it swapped an address near wallet or payment wording, and on iPhones and iPads loaded iOS exploit-kit stages.
  • The attacker's wallets had received about $7,040 by 14:25 EST on 16 November, by Report URI's count; in our reading that is everything paid in, not confirmed victim losses. Report URI says it cannot see how many cards, cookies or devices were compromised. Alex Thorn of Galaxy Digital estimates that victims sent only about $2,300, and says most of the Bitcoin looks like the attacker's own funds routed back through an exchange (Cointelegraph, reporting his estimate at 23:43 EST on 16 November); StepSecurity says it cannot confirm that. StepSecurity later recorded a further 1.7065 LTC (about $75, its conversion) arriving at about 05:18 EST on 17 November, after the attacker had swept the wallets; it was still there at 14:51 EST.

Optimizely told BleepingComputer the vulnerability was "in the Experimentation account signup flow." Its 17 November interim report adds that accounts the attacker created through that flow could "act as existing users on customer accounts through a combination of public information and brute force," letting the attacker edit Project JavaScript as administrator users. Optimizely says this "was not the result of a compromised Optimizely administrator account" and was not a credentials breach.

Facts that limit the alarm: Optimizely puts the exposure window at 02:51 to 17:24 EST on 16 November; Report URI's independent tracking saw the code from 23:21 EST on 15 November until 18:11 EST on 16 November, about 19 hours. Cards entered in a payment provider's cross-origin iframe (Report URI names Stripe and Adyen) were out of this skimmer's reach. Hudson Rock says the evidence links the iOS branch to exploit-kit delivery infrastructure, but "does not establish a zero-day, successful device exploitation, or collection of specific device data by unseen downstream stages." Hudson Rock also says the evidence does not establish a compromise of the named websites' backend systems.

For third-party risk teams, the lesson is structural: none of these sites shipped bad code; a vendor allowed to run code on every page did. The questions are which suppliers' customer-facing pages load Optimizely Web or similar tag scripts, whether they were affected, and who can change what those scripts serve.

What Happened

All times below are Eastern Standard Time (EST, UTC-5), converted from the UTC times in the cited sources. Note the date rollover: the first Optimizely injection Report URI saw was late on 15 November EST. Optimizely's interim report gives a later start (02:51 EST on 16 November) and an earlier clean-up (17:24 EST) than Report URI observed, and says its 11:06 EST Cloudflare block of malicious outbound code "based on known patterns" fully blocked delivery, while Report URI's sweep found the code in 186 snippets between 14:24 and 14:59 EST. Neither has explained the differences, so both are shown.

Time (EST) Event Source
13–16 Nov An earlier wave serves a small loader from the same injector domain the first BBC Optimizely injection loaded, through Kissmetrics, StorifyMe and Minute Media (minute-ly) scripts on unrelated sites. Report URI attributes it to the same actor; Hudson Rock says the shared infrastructure links the delivery activity but does "not prove that one person controls all affected vendors." Report URI; Hudson Rock
15 Nov, 19:37 First impersonated user account created through the signup flow. Optimizely interim report
15 Nov, 23:21 First Optimizely injection seen, in the BBC snippet. Report URI
16 Nov, 00:23 The main payload domain is registered. Report URI
16 Nov, 02:51 First malicious code written to a customer project. Optimizely interim report
16 Nov, 08:34 Optimizely detects the malicious activity. Optimizely interim report
16 Nov, 08:43 Card skimmer added to the payload. Report URI
16 Nov, 10:06 Report URI's pipeline flags the first malicious activity on a customer website, triggering its investigation. Scott Helme
16 Nov, 11:06–11:53 Optimizely blocks malicious outbound code at Cloudflare (11:06), closes the signup-flow flaw (11:10), and ends the impersonated users' sessions and revokes their tokens (11:53). Optimizely interim report
16 Nov, 12:13–12:57 A second iOS host is added; an exploit iframe is injected for visits on iOS 18.6 or earlier. Report URI
16 Nov, 14:24–14:59 Report URI's sweep finds 186 of 1,611 snippets infected. Report URI
16 Nov, ~15:40–16:10 Mass clean-up of infected snippets. Report URI
16 Nov, 16:52 One of 188 infected snippets is still infected. Report URI
16 Nov, 17:24 Optimizely removes the injected code from all affected projects. Optimizely interim report
16 Nov, 18:11 All infected Optimizely files Report URI tracked are clean. Report URI (17 Nov update)
16 Nov, 20:02 Optimizely posts its status page incident, component "Snippet Delivery". Optimizely status page
16 Nov, 23:11 – 17 Nov, 00:39 The attacker sweeps funds from the LTC, USDT and BTC wallets. Report URI (17 Nov update)
17 Nov, ~05:18 An additional 1.7065 LTC (about $75) arrives in the attacker's Litecoin wallet; StepSecurity says it was still there at 14:51 EST. StepSecurity (17 Nov, 14:55 EST update)

The payload was rebuilt 26 times during the day, according to Report URI. Report URI also says Smartsheet's snippet was cleaned at one point and then re-infected with a base64-encoded loader, "so whoever wrote the code still had access at the time." Report URI also reports the same actor's injector domain on other sites since 17 September, and a cash-out address that the attacker also used in operations in March 2026 and on 3 November 2026.

Optimizely's statement. Beyond the status page text, its statement to BleepingComputer adds that it "blocked access to the vulnerable path and deployed a fix," "ended active sessions, revoked tokens, and reset access for affected user accounts," blocked code delivery "in Cloudflare," and "notified all affected customers." An Optimizely spokesperson gave Gizmodo the same signup-flow explanation.

One affected customer's response. Salesforce's Trust Status advisory, posted at 22:29 EST on 16 November and still active without updates when we checked on 18 November, says modified Optimizely scripts on "certain Salesforce websites" could load malicious code in visitors' browsers. As a precaution, Salesforce says it "is confirming Optimizely scripts are disabled across preview and production instances," and will notify affected customers "without undue delay" if it identifies unauthorized access to their data. For third-party risk teams (Rescana analysis): as worded, that notification trigger is unauthorized access to customer data, and the advisory does not say whether data your staff or customers typed into an affected Salesforce page would count, or which sites, pages or hours were affected. If Salesforce is in your supplier inventory, record the advisory against it and ask which Salesforce-hosted pages loaded the modified scripts, between which times, and whether any were login or customer-facing pages.

What has not been disclosed:

  • The final root cause. Optimizely has described the flaw only at a high level (attacker-created accounts acting as existing administrators through public information and brute force) and says it will confirm the root cause and final indicators when Mandiant's investigation is complete.
  • Why its timeline starts later and ends earlier than Report URI's independent observations.
  • How many projects or customers were changed. Optimizely says it cannot disclose its customers and is confirming individually with each one.
  • How the linked loader got into the Kissmetrics, StorifyMe and minute-ly scripts, and whether one operator is behind both waves. Hudson Rock says each provider's initial access mechanism "remains unknown."
  • How many visitors loaded the code, and how many cards, sessions or devices were actually compromised.

Technical Information

Delivery. Report URI saw three loader variants in Project JavaScript: a plain script tag pointing at the payload host; the same URL hidden in base64 and decoded with atob(), which a plain search for the domain will miss; and a full inline copy of the payload that runs through eval if the payload host is blocked. Hudson Rock says the BBC and HSBC UAE specimens embed a complete payload that runs if the remote fetch errors or has not finished after 2.5 seconds. Snippets that customers serve from their own domains were affected as well, because the content comes from the same Optimizely project. Report URI says this "was not a CDN cache-poisoning attack": the CDN served what the project contained.

Payload capabilities (Report URI, corroborated by BleepingComputer and StepSecurity):

  • Address swapping. Replaces checksum-valid wallet addresses across eight chain families in page text, form fields, links, the clipboard and outgoing fetch/XHR requests. Sources differ on gating: Report URI describes it swapping any valid address it finds, while StepSecurity says it swaps only near payment words such as "wallet" or "deposit". A copied address can already be the attacker's when pasted.
  • Card skimming. From 08:43 EST, it watches form submissions and every fetch/XHR POST for a Luhn-valid card number, expiry and CVV, then sends them with navigator.sendBeacon. StepSecurity says it caches captured cards in sessionStorage (keys starting cc_), leaving card data in the victim's browser.
  • Cookie and screenshot theft. When it swaps an address near wallet or payment wording, it sends the page URL, document.cookie (cookies not marked HttpOnly), the original addresses and an html2canvas screenshot.
  • iOS stages. On iPhones and iPads it loads router scripts that select an exploit chain by iOS version. Report URI identifies the Coruna kit for iOS 13 to 17 and DarkSword for iOS 18.x, and says the kit's own comments list iOS 18.7.3 and later, and 26.3 and later, as patched. Hudson Rock cautions that the urlscan records it reviewed were delivery tests, not real-device exploit validations.

Why CSP and SRI mostly did not help. Subresource Integrity pins a script to a hash, but Optimizely's own documentation says the snippet "automatically updates to run any experiment you create" and tells customers not to modify it. That is incompatible with a fixed hash on the live CDN URL: every legitimate publish would break the tag. The exception is a gated self-hosted copy: pull the snippet, diff and approve it, then serve that frozen copy from your own origin, optionally pinned with SRI, so an unapproved Project JavaScript change cannot reach visitors. An auto-sync proxy that re-serves whatever Optimizely publishes, as the affected customer-domain snippets appear to have done, gives no such protection. The cost is slower experiment releases (Rescana guidance). Allow-listing cdn.optimizely.com in a Content Security Policy did not help either, because the malicious code arrived from that host. Report URI crawled 256 of the affected organisations' own hosts and evaluated their CSPs. Its results table covers 159 organisations (our sum of its four rows):

  • 68 had no CSP at all.
  • 50 had a CSP that allowed both the payload and the data theft.
  • 17 ran the payload but blocked exfiltration with connect-src. Address swapping still worked, because it happens inside the page.
  • 24 (15%) blocked the payload with a restrictive script-src that did not list the payload host.

Report URI says two common settings let the payload through: 'strict-dynamic', which trusts any script the trusted snippet inserts, and 'unsafe-eval' (or 'unsafe-inline'), which the inline fallback needs; Report URI says Optimizely's documentation tells customers to allow 'unsafe-eval'. At HSBC, it says, the CSP blocked the remote payload and the data theft, but the inline fallback still ran through eval.

Affected Product Versions

N/A, SaaS incident. There is no customer-installed Optimizely Web Experimentation version to upgrade.

  • Affected: "some customer projects" in Optimizely Web Experimentation, according to Optimizely. Report URI lists 178 organisations from the 1,611 snippets it could check; its list is a lower bound, not a census.
  • Also served from customer domains: self-hosted or proxied snippets that pull from the same Optimizely project.
  • Optimizely's PCI CDN: Optimizely says the injection "bypassed our protections hosting our high-security PCI script" (cdn-pci.optimizely.com).
  • Not affected, per Optimizely: Feature Experimentation, the SDKs and legacy Full Stack.
  • Related delivery channels, linked infrastructure: Kissmetrics, StorifyMe and minute-ly scripts from 13 November. Report URI attributes them to the same actor; Hudson Rock links them through shared loader infrastructure but not to a proven single operator, and says a response that only delivers this loader establishes the iOS router path, not the card or wallet modules. These are separate vendors and are not covered by Optimizely's statement.

Workaround and Mitigation

  1. Read every project's Project JavaScript and every experiment's custom and shared code in the Optimizely UI, not just the served file. Report URI advises looking for the attacker domain, atob( and any script your team did not write. A clean snippet today may still have served the code on 16 November.
  2. Pull the change history. Optimizely's documentation says each project's History logs all Project JavaScript changes and can be filtered by Source (REST API or the Web Experimentation UI). Review the entries for 13–17 November and match each change to a known collaborator and change ticket. Optimizely says entries attributed to "Optimizely Support" are its engineers restoring code, and that where a project already had custom Project JavaScript the attacker's code replaced it; ask Support to restore the original.
  3. Review who and what can write to the account. List collaborators, API tokens and integrations, and remove anything you do not recognise. Report URI advises rotating credentials; Optimizely says customers do not need to reset passwords or regenerate API tokens, because the attacker impersonated users rather than using their credentials. Rotating anything you cannot account for is cheap (Rescana guidance). Ask your Optimizely Customer Success Manager for your per-site timeline and application logs, which Optimizely says it can provide. SSO and MFA would not have stopped this flaw, which Optimizely says bypassed authentication, but they still close the ordinary credential path (Rescana guidance).
  4. Scope exposure by page, not by site. Determine whether checkout, login, account and crypto-deposit pages loaded the snippet from the first unauthorised change in your project history until your own verified-clean fetch. Optimizely gives 02:51 to 17:24 EST on 16 November overall, with per-customer windows from your Customer Success Manager. Report URI's tracking suggests 15 Nov 23:00 EST to 16 Nov 18:15 EST, but it saw only the snippets it could find, and Smartsheet's was re-infected after a clean-up. Until the two are reconciled, use the union of both for every lookback, including web, CDN and proxy logs and CSP violation reports: 15 Nov 23:00 EST to 16 Nov 18:15 EST, extended to your own verified-clean fetch for self-hosted or proxied copies. In our reading, Optimizely's figures come from its own logs of project changes and clean-up, while Report URI's come from snippets it saw being served; neither source explains why Report URI saw the code about three and a half hours before Optimizely's first logged write, so put that question to Optimizely (Rescana guidance). Hudson Rock advises treating financial-theft exposure and mobile-stage delivery as separate investigations.
  5. Treat card data typed into your own forms on 16 November as potentially exposed, and rotate session tokens or secrets held in non-HttpOnly cookies (Report URI). Involve your acquirer early (Rescana guidance).
  6. Preserve evidence before you clean up. Keep served response bytes, hashes, headers and project revisions (Hudson Rock). Report URI says it can share its archive of infected snippets with affected organisations and CERTs on request.
  7. Harden CSP on sensitive pages first. Restricting connect-src to named hosts is what held here: Report URI found it blocked card, cookie and screenshot exfiltration on 17 affected sites even where the payload ran. Restricting frame-src likewise keeps out the hidden exploit-kit iframe. 'strict-dynamic' loosens neither, because it applies only to scripts. In script-src, 'unsafe-eval' lets an allowed script compile strings into code, which is how the inline fallback ran at HSBC; 'unsafe-inline' lets inline script blocks run. Avoid both. 'strict-dynamic' with nonces is a sound anti-XSS design, but it makes the browser ignore host allow-lists and trust whatever an allowed script loads (MDN), so it cannot contain a compromised vendor script. If Optimizely needs 'unsafe-eval' or you rely on 'strict-dynamic', do not load Optimizely on login, checkout and account pages. Enable violation reporting with 'report-sample' (Rescana guidance, drawing on Report URI's findings). Optimizely's report says sites with a strong CSP were protected; in our view, Report URI's HSBC case shows that "strong" has to exclude 'unsafe-eval'.
  8. Take tag scripts off payment pages. Use the payment provider's hosted fields in a cross-origin iframe, which this skimmer could not read (Report URI, StepSecurity). An iframe is not immunity: a script on the parent page can still hide, overlay or replace it with a fake form (Rescana guidance), and PCI SSC still requires SAQ A merchants to "confirm their site is not susceptible to attacks from scripts that could affect the merchant's e-commerce system(s)." Under PCI DSS v4.0.1, Requirement 6.4.3 requires each payment-page script to be authorised, integrity-assured and inventoried with a written justification, and Requirement 11.6.1 requires change- and tamper-detection on security-impacting HTTP headers and script contents as received by the browser, at least weekly or at a frequency set by targeted risk analysis (PCI SSC summarises both). In our view, an auto-updating experimentation snippet is hard to square with 6.4.3, and a weekly check would very likely have missed this injection, which by Report URI's timeline was served for about 19 hours and skimmed cards for under 10. Treat compliance as the floor. Optimizely says this injection also bypassed the protections on its PCI-designated snippet host (cdn-pci.optimizely.com). A supplier that loads Optimizely on payment pages from that host should treat those pages as in scope for this incident, not as protected by the PCI designation; ask for its 6.4.3 script authorisation and 11.6.1 tamper-detection evidence for 16 November (or, for an SAQ A merchant, the basis for its confirmation that its site is not susceptible to script attacks), and whether it has informed its acquirer (Rescana guidance).
  9. Tell staff who use iPhones and iPads to update to iOS 18.7.3 or later, or iOS 26.3 or later (Report URI). Devices on older versions that reached the iOS stage hosts between 13 and 16 November (see the staff hunting under Indicators of Compromise) are candidates for triage, not presumed compromised.

Indicators of Compromise

Optimizely published preliminary indicators (an exfiltration endpoint and skimmer-hosting domains) in section 10 of its interim incident report and says they may be updated. They are not the same hosts the researchers list, so hunt for both sets. Report URI's write-up contains the indicator table (attacker domains, payload hashes, global markers and wallet addresses); see the Indicators of compromise section in Report URI's post and Hudson Rock's IOCs and hashes. We do not reproduce vendor or third-party indicators here. When you hunt:

  • Your visitors: your DNS and proxy logs will not show their browsers. Your evidence is CSP violation reports, any client-side monitoring you run, logs from a self-hosted snippet proxy, and project history. No hits in your network logs does not mean no visitor exposure (Rescana guidance).
  • Your staff: DNS and proxy logs show which employee devices reached the attacker hosts while browsing any affected site, whether or not your own sites use these vendors. Search from 13 Nov 00:00 EST for the Kissmetrics, StorifyMe and minute-ly wave and from 15 Nov 23:00 EST for Optimizely (the start of the combined window in step 4; no source reports the code being served earlier, and Optimizely's 19:37 EST attacker sign-up would not appear in your staff's logs), and feed iOS hits into step 9 (Rescana guidance based on the timelines above).
  • A domain block alone is not containment: the inline and embedded-fallback variants run without the payload host (Report URI, Hudson Rock).
  • Report URI says the main payload and earlier injector domains are on clientHold. Historical log hits still matter.

What to Do With This Vendor Today (TPRM)

Experimentation, tag and analytics scripts are a fourth-party dependency that rarely appears in a supplier review: your supplier's login page or checkout runs code that a vendor you never assessed can change at any moment, with no release on the supplier's side. Here one vendor's project store became a shared failure domain for 178 organisations' front ends by Report URI's count, and the same loader infrastructure was served through three other providers' scripts that week (Report URI attributes this to the same actor; Hudson Rock says the evidence links the delivery activity but does not prove common control, and each provider's initial access remains unknown).

Ask suppliers whose customer-facing pages handle your users, payments or data:

  1. Do you load Optimizely Web Experimentation, or Kissmetrics, StorifyMe or minute-ly scripts, on any page that handles our users, payments or data? Which pages, and from which host, including self-hosted snippets and Optimizely's PCI-hosted snippet on cdn-pci.optimizely.com?
  2. Did Optimizely notify you that a project of yours was affected? If so, which pages served the code, between which times, and what did your review of project history show?
  3. Did the code run on login, checkout, account or crypto-transfer pages? Do those pages collect card numbers in your own form fields or in a provider's hosted iframe?
  4. Who can publish code through your experimentation and tag vendors? Is console access behind SSO and MFA, are API tokens scoped and rotated, and are Project JavaScript changes reviewed like a production release?
  5. What CSP do you enforce on sensitive pages? Does it allow 'unsafe-eval' or 'strict-dynamic', do you collect violation reports, and would you have seen a new script host appear?
  6. Do you inventory client-side scripts, with an owner for each, and monitor them for change? On payment pages, PCI DSS v4.x expects each script to be authorised, integrity-assured and inventoried with a justification (6.4.3), and headers and script contents as received by the browser to be checked for tampering at least weekly or per risk analysis (11.6.1). How often do you actually check, and would you have caught a change that lasted under a day?
  7. What do your contracts with script vendors require? Change notification, exportable audit logs, incident notification timelines, and whether the vendor can version or pin what it serves.

If you use Optimizely Web Experimentation yourself, ask the same questions internally and record the answers against the vendor record. What to ask Optimizely (Rescana guidance): your per-site timeline and application logs (Optimizely says your Customer Success Manager can provide them); what data in the Experimentation Platform the impersonated administrator could see, which Optimizely says was anything that user could access; why its timeline differs from independent observations; the written third-party-risk confirmation it says it is still preparing; when Mandiant's root cause and final indicators will be published; and when Project JavaScript approval, which Optimizely lists as a feature request, will ship, along with change notifications, versioned or pinnable snippets and a mode that does not need 'unsafe-eval'; what rate limiting and alerting covered the signup flow, given that its report describes repeated guessing of employee email addresses, up to thousands of times per customer. Optimizely says it sends updates to customers individually and daily until it closes the incident; its public status page has not been updated since 20:02 EST on 16 November, so ask your Customer Success Manager rather than relying on the status page.

Need help finding which of your suppliers run Optimizely or other third-party scripts on sensitive pages, and tracking their answers? Talk to an expert

Sources

Talk to an expert

Talk to Rescana about this advisory — or about how we help TPRM teams keep reviews moving.