top of page

Hop by Hop: The Optimizely Skimmer Was the Sixth Delivery Channel in Two Months. It Cost Maybe $500 to Run, Took About $7,000, and Handed Everyone Else a Seven-Figure Bill

Writer: Patrick Duggan
Patrick Duggan
4 hours ago
16 min read
Hop by Hop: The Optimizely Skimmer Was the Sixth Delivery Channel in Two Months. It Cost Maybe $500 to Run, Took About $7,000, and Handed Everyone Else a Seven-Figure Bill

If any page you run loads a third-party script, from analytics, A/B testing, story widgets or video players, assume this crew has already tried it. The Optimizely injection that hit 178 organizations on Monday, November 16, 2026, was not a one-off breach of one vendor. Public scan records show the same operator moving through at least six delivery channels since September 13: an SEO "connector" panel on small sites, then Kissmetrics, then StorifyMe, then a script host called varors, then Optimizely, then Minute Media. Block the domains at the end of this post, give every third-party script on your pages a Content Security Policy that names the hosts it is allowed to talk to, and update every iPhone and iPad.


This is the architecture, layer by layer, from the vendor that carried the code down to the exchange account the money went to. Scott Helme at Report URI found the campaign and wrote the definitive account of the Optimizely wave; Matt Lucas at RedEye Security swept 1,246 companies and notified the exposed ones. Our first post covered the payload. This one follows the wires.





Layer 1: the delivery channels


We pulled every urlscan.io capture that touched the two injector domains and the loader, opened each one, and read which script on the victim's page made the request. That is the browser's own record of who called whom. In date order:


September 13 to November 10: the test bed. A script called "connector.js" from linkforge.cc, served with a site key, shows up on small business sites: a Turkish telecom installer, a Brazilian investment site, Russian, Azerbaijani, Ukrainian and Polish businesses, a South African group, a US church. It calls home to linkforge.cc/api/connect and usually loads alongside niko-ngx-schema.com. Report URI describes linkforge as an SEO "backlink panel": site owners install it to sell links, and its operator decides what runs. From September 17 it starts loading webshell.lol/get.js on some of those sites. Small sites, cheap traffic, nobody watching: a place to test.


November 13: Kissmetrics. On Friday, the analytics script from i.kissmetrics.com and i.kissmetrics.io starts loading webshell.lol/get.js. The first capture is Security Bank in the Philippines at 05:42 UTC, then a UK insurer, two Canadian retailers, a US marketing software company, a Japanese Google Workspace reseller and a Chicago IT firm.


November 13 to 14: StorifyMe and varors. The story-widget script from cdn.storifyme.com and cdn.storifyme.xyz loads get.js on Frankfurter Allgemeine Zeitung, the Sydney Morning Herald, the University of Basel and several vidaXL country stores. A third host, s16.varors.com, does the same on two smaller sites.


November 16, 04:21 UTC: Optimizely. The Project JavaScript field in customers' Optimizely projects starts loading the injector, first webshell.lol/get.js on bbc.co.uk and bbc.com, then from 08:28 the new loader, 1688360.cc/track.js. Report URI counts 178 organizations, including Optimizely's own website. Whether Optimizely's platform or a shared access path was breached has not been said by anyone who knows.


November 16, 16:26 UTC: Minute Media. Fox Sports loads get.js through the video player script on apv-static.minute-ly.com.


Two cautions on reading this. First, a script that wraps the browser's DOM functions shows up as the "caller" for everything after it. On the Air France, KLM and Fox News pages the recorded caller was a client-side security product (Jscrambler and Confiant), almost certainly because those products hook script insertion, not because they were compromised. Second, we checked whether minute-ly.com was a lapsed domain someone re-registered, because its registration date is January 13, 2026. Its certificate history and Report URI's account both point to Minute Media's own domain, so we dropped that theory.


The pattern is the finding: an SEO panel, an analytics library, a story widget, an A/B testing platform and a video player in nine weeks. Each one is a script that thousands of sites load without reading. The operator is not breaking into websites. It is breaking into the things websites trust.



The short version, in plain English


Big websites are built like a house with a lot of contractors. The company writes its own pages, then lets outside companies add little pieces of code: one counts visitors, one tests which button color sells more, one shows a swipeable story, one plays videos. Those pieces of code run with the same power as the company's own code. Whatever they say to do, your browser does.


These tools are bought by marketing teams, not security teams. That is why they rarely get a security review, and why one hacked contractor can reach hundreds of houses at once.


Who uses the four tools this crew rode in on, by the sites we saw in the scan records:


  • Optimizely is for testing changes to a website and personalizing it. It is used by very large brands: the BBC, CNN, Fox News, Zoom, Salesforce, GitLab, HSBC, Equifax, Air France-KLM, Ledger, Peacock, Starbucks Canada, American Girl, Scheels, Smartsheet, Stamps.com and Mutual of Omaha were all on the list. Lots of visitors, lots of logged-in customers, lots of card payments. The best target of the four.

  • Kissmetrics is for tracking what visitors do on a site. We saw it on smaller companies: a bank in the Philippines, a UK insurer, Canadian online stores, a marketing software firm and two IT companies. Smaller, but banks and insurers mean logins and money.

  • StorifyMe makes Instagram-style story widgets for websites. We saw it on news sites and stores: Frankfurter Allgemeine Zeitung, the Sydney Morning Herald, the University of Basel and vidaXL.

  • Minute Media runs a sports video player that other publishers put on their pages. We saw it on Fox Sports. Sports fans overlap heavily with betting and crypto users.

What they have in common: lots of people, many of them with money, crypto or a card in hand. The crew did not need to break into a single one of those websites. It only needed to break into the contractors they all trust.



What to do if you use one of these tools


It depends on where the bad code lived.


With Optimizely, the bad code was put inside a box you control: the Project JavaScript, custom code that runs on every page. So you can check it yourself. Open your project settings and every active experiment, delete anything nobody on your team wrote, and look in the change history for edits around November 16 that nobody can explain. Then remove people who no longer need access, turn on two-step sign-in, and change the passwords and API tokens. If you copy the Optimizely file onto your own server, delete that copy and pull a fresh one after cleaning up.


With Kissmetrics, StorifyMe and Minute Media, the bad code came from the vendor's own servers, so as far as we can tell there is no box for you to clean. Turn the tool off until the vendor tells you in writing what happened and when, or host a checked copy yourself and lock it with a Subresource Integrity hash so the browser refuses a changed file. Change your logins and keys with that vendor anyway. Search your logs back to November 13 for the domains at the end of this post.


For everyone: a Content Security Policy that lists the only places your pages may load code from and send data to would have stopped this whichever contractor carried it in.



Layer 2: the loader


1688360.cc was registered at 05:23 UTC on November 16 at Spaceship. A sibling with the digits reversed, 3601688.cc, followed at 06:27 on the same Cloudflare nameserver; VirusTotal's passive DNS shows it pointing at the same server that day. That server is 136.113.106.136, which Google's published ranges place in Google Cloud us-central1, in Iowa. Its certificates are ECDSA P-256 from Let's Encrypt plus the kind Cloudflare issues automatically at its edge.


track.js was rebuilt 26 times on November 16 according to Report URI. The build we read swaps wallet addresses in the page, the clipboard and outgoing requests, posts card numbers to card.php, posts cookies and html2canvas screenshots to hit.php, and on touch devices loads the iPhone router.



Layer 3: the iPhone stage


The router came first. 5vobauua3ojpklx.icu, je4ft3gn8vxudum.icu and p55abnw2pnsz0bo.icu were registered at 17:52 UTC on November 9 and share one Let's Encrypt certificate. They point at 34.96.213.201 in Google Cloud asia-east2, Hong Kong, behind a stock nginx. The same day, the same address answered to c2.34-96-213-201.sslip.io and mch.34-96-213-201.sslip.io, a free service that turns any IP into a hostname. The operator named its own server "c2."


The exploit kits came last. eil786npto9u5bq.icu, 1y69gy2mdzaltz1.icu, gm38duxoihbmy5o.icu and an adm. admin subdomain were registered at 16:38 UTC on November 16, certified minutes later with a single private key across all four names, and served from 34.87.36.58 in Google Cloud asia-southeast1, Singapore, on nginx 1.24 for Ubuntu. Per Report URI and RedEye, the router reads the iOS version and loads Coruna for iOS 13 to 17, DarkSword for iOS 18, and a fingerprinting page for iOS 26. Both kits are public; we wrote in October that DarkSword had reached GitHub.


Telemetry goes to l7bcsnwc7nzr3uo.icu, registered October 28 at Dynadot and kept behind Cloudflare. Of the fallback addresses hard-coded in the router, ios-hack.icu was not registered until 23:39 UTC on November 16, after the campaign was public. It was the good guys. The domain was registered through Cloudflare's registrar with a US registrant, nothing like the operator's habits, and when we had urlscan load it privately on November 17, the home page and the /boot.js path both served a page titled "Sinkholed by Censys." Censys took the fallback address away before the operator could use it. Any iPhone that tries it now reaches a sinkhole instead of an exploit kit, and a lookup of it in your logs still means a device ran the router.



Layer 4: registration, DNS and keys


The .icu names came through Global Asset Domains with the registrant country given as CN. Their nameservers, share-dns.com and share-dns.net, were registered in 2022 through Gname, a Singapore registrar, with registrant country SG. Everything else used cheap, privacy-masked registrars: Spaceship, Namecheap, Hostinger, Dynadot and Cloudflare. Every injector and the loader sat behind Cloudflare.


Every certificate in the campaign is ECDSA P-256, the default of the common free certificate tools; that is a habit, not a fingerprint of a person. The one useful key finding is inside the campaign: the four kit names share one private key, so they are one machine stood up by one script. Our attempt to find the same keys on any outside domain through the crt.sh public-key search returned nothing for every key, including ones we know exist, so that check is unknown, not negative.



The family it belongs to


The iPhone kits were not new to us. Our corpus already held 255 Coruna indicators, from the ThreatFox and URLhaus feeds since September 5, and 81 DarkSword indicators, including our own research since April. When we laid this campaign's infrastructure next to them, three links showed up:


  • Same DNS provider. chainmate[.]top, already tagged DarkSword in our corpus, uses the same share-dns.com and share-dns.net nameservers as this campaign's iPhone stage.

  • Same naming habit. Our Coruna set includes random 15-character .icu names, such as aygjqdr4l2s88zi[.]icu and go32ip.74rcu0ntwgvtpzn[.]icu, exactly the shape of 5vobauua3ojpklx[.]icu. It also includes lucky-number pinyin like faicai888, 发财, "get rich," next to this campaign's JIUJIU888.

  • Shared hosts between the two kits. fc.rsqqq[.]top, st.onlinefc[.]top and 66ds[.]lol appear under both the Coruna and the DarkSword labels in our data.

What does "same naming habit" mean, explained like you're five? When a computer makes up a website name on purpose to look like nonsense, the person running it still has to pick the settings: how long, which letters, which ending. This crew keeps picking the same ones: 15 characters, a jumble of letters and numbers, ending in .icu. It is like handwriting. Two notes in the same handwriting are not proof the same person wrote them, but it is the first thing a detective checks. The lucky numbers are the doodle in the margin. In Chinese, 8 sounds like "prosper," so 888 means "lots of luck," and faicai is 发财, "get rich." Names like faicai888 in the Coruna set and JIUJIU888 in this campaign are a superstition written into the infrastructure, the same way someone always picks 7 for their locker.


What JIUJIU888 most likely says: pinyin drops the tones, so "jiu" on its own could be several Chinese words: 九 nine, 久 long-lasting, 酒 liquor, 旧 old, 舅 uncle. Doubled and set next to 888, the two readings that fit are 九九 "nine nine" and 久久 "forever," which sound almost the same and are used as a pun on each other for luck and longevity. So JIUJIU888 reads roughly as "lasting luck, triple fortune." The same lucky-number naming is common on Chinese-language gambling and gray-market sites, and that is consistent with another marker in the code, DAILI, 代理, which means "agent" or "proxy" and is the word those businesses use for their affiliate resellers. The "data-channel-code" attribute on the iPhone loader looks like an affiliate tracking number in the same style. That is a reading of the words, not an identification of anyone. Neither clue alone names anyone. Together, on the same DNS provider, loading the same exploit kits, they say "same shop."



Where the iPhone stage came from: a public kit on GitHub


We searched GitHub's code for the unusual strings inside this campaign's iPhone loader, the configuration name __DAILI_CFG, the attribute data-daili-gate, the marker obChTK and the server paths /api/gs_state and /api/pe_ok. They led to one repository, jianbuzhai698-ui/ios, titled "神也自卑 iOS — iOS 全版本安全研究框架," roughly "God-Is-Humble iOS: an all-versions iOS security research framework." It was created on August 12, 2026, and holds 1,052 files. Despite the word "research," it is a ready-to-deploy kit: an admin panel with a login page, an upload page and a live log viewer, "beacon" and "harvest" scripts, a deployment tutorial, Coruna stages for iOS 13 to 17 and chains for iOS 18. It carries every one of those markers.


It does not contain this campaign's domains, servers or wallets, so whoever wrote it is not necessarily who ran the Optimizely wave. The kit is the shared tool. It has been forked 50 times since August, most recently on November 15, the day before the Optimizely injection. We read it as text through GitHub's API and ran nothing. We have reported it to GitHub.


A Chinese researcher has been tracking DarkSword delivery sites publicly since October 29 in ValueMovedHere/DarkSword-site-report, including a decoded loader with the same obChTK marker. Two of the eleven domains listed there, hsylc888[.]com, another lucky-888 name, and ymxim-shd[.]com, use the same share-dns nameservers as this campaign's iPhone stage. Seven more sit on a single subnet, 155.102.130.0/24, behind the same atrustdns nameservers: a second cluster. None of the eleven were in our corpus. All eleven are now in our feed at confidence 70, credited to the tracker, with a note that some may be compromised legitimate sites.


So this is one ecosystem, not three: the people running Coruna and DarkSword infrastructure and the people who planted the Optimizely loader build their domains the same way, on the same DNS provider.


Here is the lesson for anyone doing this kind of hunting, us included. None of those links is a shared IP address, a shared subnet or a shared parent domain, which is what most correlation tools, ours among them, connect on. They are habits: which nameserver provider, what a random name looks like, which kit gets loaded. A brand-new domain on a fresh cloud address has no network neighbors, so a network-only graph shrugs. The router domains were registered on November 9, a week before the Optimizely wave. Pivot on the nameserver and the naming pattern as well as the address, and that week is visible.



Where it lives: the regional picture


Pulling the geography together from WHOIS, Google's IP ranges and our own enrichment:


  • Build and operate: Hong Kong and Singapore Google Cloud regions for the iPhone stage, Iowa for the loader, a Chinese registrant country on the .icu names, a Singapore-registered DNS provider, and pinyin in the code (JIUJIU888, BIGNIAO, and DAILI, which reads as 代理, "agent" or "proxy").

  • Test: small business sites in Turkey, the Caucasus, Eastern Europe, Russia, Brazil and South Africa, through the SEO panel.

  • Harvest: big Western and Asia-Pacific brands with crypto-holding, card-carrying audiences: the US, UK, Germany, Australia, Switzerland, France and the Netherlands, the Philippines, Singapore and Japan.

That reads as a Chinese-speaking, financially motivated crew using cheap global infrastructure, not a state program. It is not North Korea: the haul is small, the cash-out is direct, and none of the laundering tradecraft is there. It is not Russian either; the Russian sites were victims in the test bed. Hosting country and registrant country are not attribution, and we are not attributing this to anyone.



Layer 5: the money


The nine replacement wallets cover Bitcoin, Ethereum and other EVM chains, TRON, TON, Solana, Litecoin, Bitcoin Cash and Dogecoin. Report URI puts the total received at about $7,040.


The Bitcoin wallet took five payments on November 16 between 11:13 and 19:38 UTC, 0.0853 BTC in total, and has never sent anything. Three of the five arrived in batched withdrawals of the kind exchanges send, which is what a victim withdrawing to a swapped address looks like.


The Ethereum wallet is the interesting one. On November 12, four days before the Optimizely wave, it received 4,610.72 USDC from a busy, unlabeled address that behaves like an exchange hot wallet. On November 14 it sent 299.9 USDC to a deposit-style address that forwarded straight back to that same hot wallet. At 13:03 UTC on November 16 it swept 4,310.82 USDC to another busy, unlabeled address. Money in from one exchange-like wallet and back out to it is what an account at that exchange looks like, and an exchange account is the point in this chain where a real name usually sits.


And the attacker is being robbed in turn: other scammers are sending its Ethereum wallet fake "USDC" tokens from look-alike addresses, the address-poisoning trick that catches people who copy an address from their own history.



The bill: attack versus defense


These are estimates. Here is the basis for each, so you can change the inputs.


The attacker's costs:


  • About 16 domains: the .icu names at a dollar or two each, the .cc, .lol and .com names at about $10 each. About $80.

  • Three small cloud servers for a few weeks: about $15 to $50, or nothing on a new-account free credit.

  • Certificates and DNS: free.

  • Exploit kits: public, so free.

  • Access to the vendors' consoles: unknown. If it came from infostealer logs, those sell for tens of dollars each.

  • Total: roughly $100 to $500.

The attacker's take: about $7,040 in crypto by Report URI's count, plus an unknown number of card numbers and session cookies. Barely steak money for a two-month campaign, which tells you the cards and cookies may be the real product.


The defenders' costs:


  • Triage at 178 organizations: even 8 to 40 hours of security staff at about $150 an hour is $1,200 to $6,000 each, or about $0.2 million to $1.1 million across all of them.

  • Card forensics: any affected site that takes cards on its own page may need a PCI forensic investigation, commonly tens of thousands of dollars each. If ten need one at $20,000 to $100,000, that is another $0.2 million to $1 million.

  • Then card reissues by the banks, session resets, customer notices and every iPhone that now needs updating, none of which we can count.

  • Total: roughly $0.4 million to $2 million, and likely more.

That is a ratio of somewhere around a thousand to ten thousand to one against the defenders. The cheapest defense on the board costs nothing: Report URI found that a strict script policy is what protected the sites that survived, and this campaign's indicators are in our free feed.


For the record, our own cost of covering it was a few hours and our standing platform, which runs about $500 to $600 a month in total.



The bigger bill: what the kits did to the economics


The exploit kits are why this crew's math works at all. DarkSword chains six iPhone vulnerabilities, three of them zero-days, and until March it belonged to commercial spyware vendors and government customers. Public exploit-broker price lists have offered several million dollars for an iPhone chain of that kind. On March 23, 2026, it leaked; we wrote that day that it could reach about 270 million iPhones running iOS 18.4 through 18.7, and by Easter it was on GitHub. Coruna covers the older iPhones, iOS 13 to 17. So the most expensive part of this attack, a working way into an iPhone from a web page, cost this crew nothing. A multimillion-dollar tool became a free download, and a crew whose whole operation costs a few hundred dollars now points it at the visitors of 178 organizations. That is the real asymmetry: the attacker's bill fell by a factor of thousands, and the defender's bill did not move.



How long our feed has covered this family


This is coverage of the family, not detection of this wave: the domains used on November 16 were new, so older indicators would not have matched them. What a subscriber had in hand:


  • DarkSword: in our corpus since April 6, 2026, as a tracked tool, and with network indicators since October 12, when we ingested the October funnel campaign that iVerify, Report URI and others published, an expired analytics domain turning old store code into an iPhone exploit funnel. That is 224 days of tracking the tool and 35 days of its infrastructure in the feed before the Optimizely wave.

  • Coruna: carried in our feed since September 5, 2026, from ThreatFox and URLhaus submissions. That is 72 days. The credit for those listings belongs to the people who submitted them; our part is that we carried them.

  • This wave: the first seven domains and three servers were in our feed at 21:23 UTC on November 16, about half an hour after Scott Helme posted them, and the rest followed through the night as Report URI, RedEye and our own pivots found them. 29 indicators in all.

And the one we did not have: the router domains registered on November 9. They shared a nameserver provider with a DarkSword domain we already held, and they used the Coruna naming habit. We were not looking for those two links. We are building that into our correlation now.



What to do


  1. Block every domain below at DNS, proxy and secure web gateway, and search logs back to September 13 for the injectors and November 9 for the iPhone stage.

  2. Inventory every third-party script on your pages: analytics, testing, widgets, video, ads, SEO plugins. For each, write down who can change what it serves.

  3. Ship a Content Security Policy that names the script hosts and the connection destinations you allow, without 'unsafe-eval', and turn on reporting so you see the next one.

  4. In Optimizely: review Project JavaScript, experiment code and change history, remove what you did not write, and rotate credentials and API tokens. In Kissmetrics, StorifyMe or Minute Media: the bad code came from the vendor's side, so disable or pin the script until the vendor explains, and rotate your credentials with them. Details are in the plain-English section above.

  5. Update iPhones and iPads to iOS 18.7.3 or later, or 26.3 or later.

  6. If card entry happened on an affected page, treat those cards as exposed; if logins did, expire sessions.


Indicators


All of these are in the DugganUSA feed under source manual-batch-helme-js-host-breach-2026-11-16, each credited to who found it.


Injectors and panels: webshell[.]lol, linkforge[.]cc, niko-ngx-schema[.]com (ours, confidence 70), varors[.]com and s16.varors[.]com (ours, confidence 70).


Loader: 1688360[.]cc, 3601688[.]cc (ours).


iPhone stage and C2: 5vobauua3ojpklx[.]icu, je4ft3gn8vxudum[.]icu, p55abnw2pnsz0bo[.]icu, eil786npto9u5bq[.]icu, 1y69gy2mdzaltz1[.]icu, gm38duxoihbmy5o[.]icu, l7bcsnwc7nzr3uo[.]icu, q1ls1ho9x7wsxlu[.]icu, ios-hack[.]icu.


Servers (shared cloud addresses, confidence 65): 136.113.106.136, 34.96.213.201, 34.87.36.58.


File hashes (SHA-256): track.js f7b84b9a94f749f8ced3f0112d0a66894e8bbc3b4aa5592021d1b8976a9bad40 and 814a963c0da480e2449d9b6664480156e1cd6d731014117401bf635badec3a0f; boot.js 8d69f32fff37208e52dc573111e8db99f031f4913845a24fb2e14755747ac860 and e410b6509f33defce16ad7896cbe2fb3cc689f2cc53d87c87d75efba6b57e290; iOS stage pages 792deb03483ce9b212df6f46319a78ea6fc39bfc5b2719c7e8b87d084524437d and 9039ebbdd6abe36ad8ac48097dae6813d4a49315e1403c9ae7ebab539601cde6 (RedEye).


We left the vendors' own domains (Optimizely, Kissmetrics, StorifyMe, Minute Media) and the exchange addresses out of the feed. They were carriers and bystanders, and blocking them would break legitimate sites and customers.



What is not known


How the operator got write access to each vendor's scripts. Whether varors and niko-ngx-schema are attacker infrastructure or other compromised services. How many cards and session cookies were taken. Whether any vendor beyond these six is serving the injector today.



How we know


urlscan.io captures and stored responses, read with our key, including the initiator of every request to the injector and loader domains. Certificate Transparency through certspotter. VirusTotal passive DNS for the three servers. WHOIS and RDAP. Google Cloud's published IP ranges and RIPEstat for regions and routing. The public Bitcoin and Ethereum ledgers through blockchain.info, BlockCypher and Blockscout. Report URI's and RedEye Security's write-ups. We did not connect to any attacker server.




How do AI models see YOUR brand?

AIPM has audited 250+ domains. 15 seconds. Free while still in beta.



Was this useful? Thirty seconds, no cookies, no tracking, no third parties, your address hashed and never stored. The question lives here: https://analytics.dugganusa.com/nps.html#post=optimizely-skimmer-hop-by-hop


Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page