⌘ K
Partner with us
Insights
All insightsResourcesAboutTalk to usPartner with us

Your Website Changed on Monday. Nobody Deployed Anything.

Understand how tag managers and experimentation tools can change live websites without traditional deployments or IT oversight.

By Editorial Team

7 min read

Your Website Changed on Monday. Nobody Deployed Anything.
DIGITAL-BUSINESS · TAG-MANAGEMENT

On November 16, visitors to 178 organizations' websites were served a card skimmer and a crypto-address swapper. None of those companies had shipped a release. The code arrived through a marketing experimentation tool, which is a second route to production that most technology leaders have never been asked to approve.


The victim list reads like a conference sponsor wall: Zoom, the BBC, CNN, Salesforce, GitLab, IKEA, HSBC, Equifax. Security researcher Scott Helme of Report URI, who found the campaign, checked 1,611 Optimizely snippets and found 188 of them carrying malicious JavaScript, spread across 178 organizations. Banks and financial firms were the largest group, with 32.

Optimizely has confirmed the cause. An attacker exploited a flaw in the signup flow of its Experimentation product, impersonated users, and typed code into a field called Project JavaScript inside customers' projects. Optimizely's content delivery network then did its job. As the security firm Strobes put it, the network "served exactly what it was given."

Most coverage has filed this under supply-chain security. That is accurate, and it lets the wrong people off the hook. The more useful reading is about how a modern digital business is organized.

The second pipeline

Every company on that list has a release process. Code is reviewed, scanned, built, staged and rolled out, with a named owner and a way back. Engineering leaders have spent a decade hardening it.

Alongside it sits another route to the same production pages. A/B testing tools, tag managers, personalization engines, chat widgets and analytics tags all work the same way: the site loads a script from the vendor, and whoever has access to the vendor's console decides what that script does. A marketer changes a headline test at 2 p.m. and every visitor's browser runs new code at 2:01. Nothing passes through the pipeline. That speed is the product. It is why these tools were bought.

Seen from the browser, the two routes are equal. Seen from the org chart, one is governed by engineering and the other by whoever holds a login to a software-as-a-service console, authenticated by that vendor's account system. On Monday the account system was the hole.

An experimentation platform is remote code execution with a friendly interface. That is a description of the feature, and it is why the feature sells.

Figure 1

Question

Release pipeline

Tag and experiment layer

Who can publish

Engineers with repository and deploy rights

Anyone with a role in the vendor console, plus the vendor itself

How the publisher is verified

Company identity provider, usually with hardware-backed sign-in

The vendor's own account system

Review before it reaches visitors

Code review, automated scanning, staging

Usually none outside the tool

Time from change to visitor

Minutes to days

Seconds. On November 16 the payload carried a 15-second cache instruction

Who notices an unplanned change

Monitoring tied to each deploy

On November 16, an outside researcher

Way back

Rollback to a known version

Remove the code and hope the vendor's history is intact

Two routes to the same production page. BusinessInfomatics editorial assessment of typical arrangements, with November 16 details from Strobes and Help Net Security. Individual companies will differ.

What sixteen hours looked like

The timeline Strobes assembled from Report URI's data, in UTC, is worth reading as an operations record. The first injected code appeared in the BBC's snippet at 04:21. A card skimmer was added to the payload at 13:43. Mass cleanup began at 20:40. Optimizely's incident notice went up at 01:02 the next morning, more than twenty hours after the first change. The company says the first malicious write happened at 07:51, and analyst Frank Dickson told CSO Online the two accounts need reconciling.

Through that day the attacker rebuilt the payload 26 times, according to Help Net Security, adding features as the hours passed. It swapped cryptocurrency wallet addresses, captured card numbers typed into a site's own forms, took cookies and screenshots, and tried to deliver exploit kits to older iPhones. The known haul is small: about $7,150 reached the attacker's wallets, and nobody can say how many cards were read.

The standard browser defense did less than its reputation suggests. Helme's analysis found only 24 of the organizations had a Content Security Policy strict enough to stop the attack, and 68 had none. Listing Optimizely as a trusted source offered no protection, because the malware came from Optimizely. The vendor's own setup guidance asks customers to permit a setting called unsafe-eval, and that allowance let a backup copy of the code run on HSBC's site.

One design choice did hold. Card details entered into a payment provider's hosted frame were out of the skimmer's reach. The companies that had kept the payment field off their own page lost nothing there.

Who signed off on the second pipeline

In most enterprises, nobody did. The experimentation tool was bought by marketing or product, added to the site template once, and has been publishing ever since. Security reviewed the vendor at purchase. Engineering approved a single script tag. From then on, the question of what runs on the checkout page on a given afternoon has been settled inside a console that IT cannot see.

Helme's summary was that "trusted JavaScript doesn't always remain trustworthy." Trust was granted to a file location years ago. What the location serves is decided fresh every few seconds.

Optimizely was the route this week. Researchers at Hudson Rock have traced similar poisoned scripts to three smaller services, without establishing that the same operator was behind them. The category is the exposure.

Five questions for every script publisher on your site

  1. Who can change it? List the people, tokens and integrations with write access in the vendor's console, and who removes them when staff leave.

  2. Where does it run? Name the pages. Then ask why any of them is a sign-in, account or payment page.

  3. How would we know? Identify what alerts when the script's content changes between releases.

  4. What can it reach? Check whether your policy limits where page data can be sent, and whether it still permits unsafe-eval.

  5. What did the vendor promise? Find the notification deadline in the contract. On Monday the public notice trailed the first change by most of a day.

What this means for technology leaders

Count publishers, not scripts. A script inventory goes stale by the next campaign. A list of who can publish to production, through which console, is short and stable, and it is the list an auditor will eventually ask for.

Draw a line around the pages where people type secrets. Experimentation on a pricing page is a business choice. Experimentation code on the page that takes a card number or a password is a risk decision, and it should be made by someone who knows they are making it. Server-side testing exists for those flows and passes through the normal release process.

Move the browser policy from report to enforce. Dickson's advice to CSO Online was plain: restrict which hosts scripts may load from, restrict where data may be sent, and drop unsafe-eval wherever the tools allow. Monday's numbers show how few companies have finished that work.

Then settle ownership. The fastest way to change your website is also the least supervised. The marketing team did not create that arrangement, and it is unreasonable to expect them to secure it alone.


Sources and notes. Help Net Security, "Malware spread through Optimizely hits 178 organizations, including BBC, CNN and Zoom" (November 17, 2026): Optimizely's confirmation, victim counts and sectors, payload behavior, 26 rebuilds, wallet receipts, Content Security Policy findings, related services traced by Hudson Rock, Scott Helme's statement. Evan Schuman, CSO Online (November 17, 2026): the signup-flow flaw as Optimizely described it, the 07:51 and 04:21 discrepancy, Frank Dickson's comments and policy advice, hosted payment frames. Strobes, "Optimizely hack hit 178 organizations without a single deploy" (November 2026): UTC timeline, 15-second cache header, sector counts, remediation steps. Report URI's own write-up and Optimizely's status notice were not read directly; figures are as those outlets reported them. Strobes sells security services. The policy counts published add up to fewer than 178 organizations, a gap the sources do not explain. Figure 1 is editorial assessment. Journalism, not security or procurement advice. Corrections welcome.

Tagged

#digital-business#tag-management#website-governance#production-security#optimizely