Optimizely confirms that an attacker exploited a vulnerability to add malicious JavaScript to customer experimentation projects. Its initial notice, posted November 16 at 8:02 p.m. EST, says the vulnerability is fixed and the injected code removed. Its subsequent report supplies the review window and customer actions below.
Scott Helme’s investigation for Report URI found 188 infected snippets associated with 178 organizations, among 1,611 snippets examined. These are script and organization counts, not counts of infected devices or people whose data was stolen.
A trusted script carried an untrusted change
Optimizely Web lets customers test changes to their websites. Its delivered script includes Project JavaScript, custom code that runs on pages loading the project. Helme found the attacker’s code in that field, while the rest of the snippets he examined remained intact.
The browser received the modified file from Optimizely’s trusted delivery domain. Some customers served the same project content through their own domains. Changing the delivery hostname alone would not have removed the injected project code.
Lawrence Abrams at BleepingComputer independently reviewed modified files and recorded website scans. He confirmed malicious scripts loaded, or attempted to load, on sites including Zoom, BBC, CNN and GitLab. That evidence concerns affected webpages; it does not establish compromise of those companies’ desktop applications or every service they operate.
Evidence status: vendor-confirmed exploitation and malicious code delivery, supported by researchers’ script analysis. The initial status notice confirms a vulnerability without identifying a CVE or affected software version. The more detailed vendor statement below narrows the entry point.
Update — November 17, 2026, 12:53 p.m. EST
BleepingComputer quotes a fuller Optimizely statement identifying the Experimentation account signup flow as the vulnerable path. The company says it blocked that path, deployed a fix and removed injected code. It also ended active sessions, revoked tokens and reset access for affected user accounts, and says it notified all affected customers.
This adds a specific entry path and account-containment actions to the earlier, less detailed status notice. The provider’s access reset cannot recall card details or cookies copied from an affected webpage. The later report clarifies that customers do not need to regenerate Optimizely API tokens solely because of this incident.
Update — November 18, 2026, 10:55 a.m. EST
Optimizely’s November 17 initial customer report, linked by Brian Krebs’s subsequent reporting, changes the response guidance. The earlier statement here that a customer checklist was unavailable is now outdated.
The vendor says the signup flaw let an attacker impersonate customer administrators without stealing or using their passwords. It limits the incident to Web Experimentation, excluding Feature Experimentation, SDKs and legacy Full Stack. It reports these November 16 milestones, all in UTC:
- 07:51: first malicious project-code write.
- 13:34: activity detected.
- 16:06: malicious outbound code blocked at Cloudflare.
- 16:10: signup vulnerability closed.
- 22:24: injected code removed from all affected projects.
Optimizely asks customers to review 07:51–22:24 UTC and request their exact per-site window and project records through support. Preserve web, delivery-network, browser-policy and payment logs. Assess pages collecting cards in their own forms, review suspicious outbound browser requests, and determine any notification obligations.
The vendor says customer passwords and Optimizely API tokens do not need resetting solely for this incident. That is separate from assessing visitor sessions readable by malicious webpage code. It has withdrawn its earlier advice to disable Web Experimentation and says customers may re-enable it. It still recommends removing unnecessary snippets from sensitive-input pages where no experiment is running.
The report also says some custom Project JavaScript was replaced, not merely appended to. Verify restoration with support. Mandiant’s investigation remains ongoing, and the vendor cannot determine successful visitor data theft merely from a script-load record.
Exposure depends on the page and device
The analyzed payload could capture card numbers, expiration dates and security codes entered into a site’s own forms. Helme says hosted payment iframes were outside that skimmer’s reach. The code also supported collecting browser-readable cookies and screenshots, with collection conditional on its payment-address replacement logic.
Helme identified delivery of Coruna and DarkSword iOS exploit kits. Delivery or attempted loading is separate from successful device exploitation. Neither reviewed investigation measures how many cards, cookies or devices were compromised.
An infected snippet also does not prove every associated webpage executed the payload. Browser policy, device version, the particular snippet and when it was loaded all affect exposure.
Review both the project and its downstream access
For operators using Optimizely Web:
- Preserve incident evidence and inspect project changes. Review Project JavaScript, experiment custom code and who could modify them. Helme observed several loader variants, including encoded and inline code; a single-domain text search is insufficient.
- Review users, permissions and integrations. Follow the vendor checklist above. Its report says Optimizely passwords and API tokens do not require blanket replacement for this incident.
- Assess payment exposure on November 16. Determine whether affected pages collected cards in their own forms. Coordinate customer communication and payment response with the relevant providers.
- Assess potentially exposed visitor sessions or secrets. Invalidate access where the evidence warrants it. The analyzed cookie-reading path could not read cookies marked
HttpOnly. - Review Content Security Policy, the browser’s resource-loading rules. Restrict script, connection and frame destinations where practical. Report-only policies observe violations; enforcement policies can block them. A trusted script that can load arbitrary descendants or evaluate injected code weakens that boundary.
Visitors who entered card details on an affected page should review charges and contact their card issuer about suspicious activity. iPhone and iPad users should install the latest security update supported by their device. The reports do not establish that every visitor needs a device reset.
What the thread adds
In the Hacker News discussion, goalieca describes frequently changing marketing trackers as an obstacle to maintaining restrictive browser policies. mcoliver proposes reviewing and deliberately releasing vendor scripts instead of accepting every remote change immediately.
Those comments identify an ownership problem: someone must approve third-party code changes and maintain the policy when marketing integrations change. Serving a vendor file locally helps only if its contents and updates are controlled. These are defensive proposals, not findings about Optimizely’s internal release process.
The replies also distinguish approving a file from trusting everything it can load. stavros asks whether an integrity check “doesn’t ensure that script doesn’t load other scripts”. Subresource Integrity, a browser check against an expected file hash, verifies the particular resource carrying that check. Content Security Policy separately restricts further resource loads. An approved loader can still be dangerous if its policy allows arbitrary scripts or evaluated code.
technion identifies the operational conflict: “the whole point of optimizely and similar tools it that it’s expected the script will change.” Pinning an approved hash means legitimate changes also need approval. Groxx warns that automatically approving every new hash would recreate the original trust problem. The proposed defense therefore needs a controlled update process alongside the browser setting.
From a 273-comment thread, in HN’s own order.
Caveat: Report URI sells script monitoring and browser-policy services, which its article promotes. Its technical findings and BleepingComputer’s checks support this capture; its product claims and protection percentages are not independent efficacy evidence.