- Amirul Cyber reports attacker code inside Optimizely bundles and other scripts on 15 captured page loads. It swaps crypto wallet addresses, skims card forms and copies cookies.
- Hudson Rock's review says this shows a multi-customer delivery problem. It does not show how the attackers got in, or how many visitors lost data.
- Check your served Optimizely bundle for the reported project ID, and monitor Content-Security-Policy violations. Two affected sites blocked the payload this way.
A clean website can still rob its visitors
Most security budgets protect what a company runs. This incident is about what a company only borrows. Amirul Cyber's researchers report that attacker code sits inside Optimizely scripts, which many sites load to run page experiments. The code targets the people browsing, not the company's servers.
In the researchers' framing, a company's own servers can be free of the problem while its visitors are still exposed. The compromise sits in the tag layer. That is the set of third-party scripts a page loads when it opens.
What the researchers found
Amirul Cyber's OptimJack research lists 15 captured page loads that loaded the injected code. Among the sites named are CNN, Zoom, Fox News, BBC, Salesforce, GitLab, American Eagle and Hootsuite. Seven more appear in the report.
Most of these loaded the code through Optimizely bundles. Three did not. Amirul Cyber's table says Fox News loaded it through an ad wrapper, American Eagle through its own first-party bundle, and an Air France page through a Jscrambler-protected script.
Hudson Rock later refined the Fox News entry. It says Fox News now has specific historical Optimizely response evidence, in a file named /js/5728063363743744.js. It adds that this finding does not depend on a Confiant wrapper.
The researchers say GitLab's project ID is 1688360. Ten more project or snippet IDs, at unrelated organisations, carry the same line.
How the attack works
The injected code does four jobs. First, a clipper. It watches for crypto wallet addresses on the page and swaps them for the attacker's. The researchers say it validates real address checksums for eight coins in all, including Ethereum, Bitcoin, Solana and Dogecoin. A pasted address can therefore change before the user submits it.
Second, a data copier. It sends the page URL, cookies and a screenshot of the page to the attacker's server.
Third, a card skimmer. It spots card numbers, expiry dates and security codes in forms and in the data a page sends on checkout. It then sends them out quietly with the browser's beacon function.
Fourth, a step aimed at Apple devices. On iPhones and iPads, the code loads a further script, boot.js, from two .icu domains. On iOS 18.6 or earlier it also adds a hidden frame. The researchers say boot.js is obfuscated and still under analysis, so its purpose is not established.
Why does this work? The researchers note that Optimizely experiment code runs with the page's own privileges. A poisoned tag therefore sees whatever the page sees.
What the evidence does not show
Hudson Rock, via infostealers.com, reviewed the research. Its analysis rests on response data archived by a researcher it names as Kirk. Hudson Rock says it matched those file hashes to the archive and recovered nine poisoned vendor files itself.
It also adds limits. The evidence shows a multi-customer delivery problem. It does not show whether attackers got in through the provider's systems, publishing credentials, integrations or separate customer accounts.
Nor does it show a breach of the named sites' backend systems. Hudson Rock says the evidence falls short of proving that each visitor received the malicious version, or that any visitor lost data. It also says audience figures for these sites must not be treated as a victim count.
Hudson Rock found that later copies of several files no longer held the injection. It says that does not cancel the earlier archived evidence.
The lesson: you ship what you assemble
A modern page is mostly assembled, not written. Each third-party script is a standing invitation to run code in your customer's browser. This case shows the cost of that trust. Your own audit can pass while the layer you do not control does the damage.
The report also shows that one control worked. Two affected sites blocked the payload with a Content-Security-Policy, a browser rule listing which hosts may supply scripts. The snippet was compromised on those sites, but the payload did not execute.
Questions to put to your team
Ask whether you load Optimizely, and which project IDs. Have someone fetch the served bundle and search it for 1688360, as the researchers suggest. Hudson Rock warns that code varies between customers, so also search for unknown .cc and .icu hosts.
Ask whether script-src violations are monitored, and who gets the alert. Ask whether checkout pages send data to hosts nobody approved. The researchers also propose a canary test: paste a dummy wallet address on a staging page, submit it, and compare what was sent.
Finally, ask who owns each third-party script's risk. If the answer is nobody, the tag layer is where an attacker will look.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Amirul Cyber.





