Skip to content
senn-techsenn-tech
Back to blog
Security
Security and data protection2026-11-17· By Franz Senn

Optimizely supply-chain attack: where a CSP stops the loader and where it does not

On 16 November 2026 an attacker put malicious code into Optimizely Web Experimentation snippets. It ran from an address many site operators had approved in their own policies, captured card and login data and swapped cryptocurrency payment addresses inside the page text. An earlier post measured the basic equipment on our sites, CSP and SRI. This one takes the question behind it, on our three domains, one real snippet and the registry of the malicious domains.

A supply-chain attack on a website hits code the operator does not write: a service provider delivers a script, the page loads it from the provider’s address and runs it with the page’s own rights. If an attacker changes that script at the provider, the malicious code runs for every visitor although the site itself was never breached. A Content Security Policy limits load sources and target addresses, but it does not judge the content of an approved script.

What the vendor published, and what is missing

The status page entry "Web Experimentation Security Incident" carries 17 November 2026, 01:02 UTC, fetched from an outside vantage point; our own resolver did not reach that page. Wording: an unauthorised party exploited a vulnerability in Web Experimentation to insert malicious JavaScript into the projects of some customers, weakness closed, code removed, customers contacted. At our check around 11:45 UTC the entry still stood on "Monitoring", no closure recorded.

Three things were absent on 17 November: no CVE, no postmortem, no count of affected projects. We treat the entry as a vendor statement. The vishing incident Optimizely disclosed on 23 February 2026 (social engineering, not a software flaw according to Optimizely) is a different case. Measurement happened independently: Report URI on the CSP headers of affected organisations, StepSecurity on scan records, DNS and wallets, Hudson Rock on recovered files. All three sell security data, which does not weaken their findings but marks their role.

More came from the vendor later. On 17 November Optimizely published an initial report for customers (nine pages, Mandiant as forensic examiner, final report pending). According to it, the attacker used a flaw in the sign-up flow to make new accounts act as existing administrators (by guessing addresses and using public identifiers, without stolen passwords) and wrote code into Project JavaScript from 07:51 UTC on 16 November. Detected at 13:34, flaw closed at 16:10, everything removed by 22:24 UTC. The status page added on 18 November, 14:49 UTC, that fewer than 5 percent of customers were affected; it still names no CVE. Two points stay open. The vendor lists pay-cloud[.]online and a workers[.]dev exfiltration endpoint as malicious domains, while the reports from Report URI, StepSecurity and Hudson Rock name 1688360[.]cc; each set of names is absent from the other source. And Report URI sees the first injection at 04:21 UTC (a BBC snippet with webshell[.]lol), while the vendor dates the first malicious code to 07:51 UTC. The vendor calls its own findings preliminary, and the conflict stays open.

Three ways the payload got in

The attacker wrote into a field customers edit themselves: the project JavaScript, delivered as cdn.optimizely.com/js/<project-id>.js. Report URI saw three forms, and StepSecurity repeats them. A tag pointing at 1688360[.]cc/track.js. The same address as a Base64 block decoded with atob(). And a complete inline copy of the loader with an eval branch. Hudson Rock found an embedded loader copy in Base64, 24,392 bytes long, in the samples from BBC and HSBC UAE; that loader tries the remote track.js first and takes the embedded copy on an error or after 2.5 seconds. Base64 changes the text inside the snippet, not the address the browser checks. Searching for 1688360 therefore catches the first form only, because the inline copy sits in the project text on the victim's own domain.

How the injected script reached the pageWeakness in thesign-up flowvendor statement, no CVEWrite into theproject fieldcustomers edit this filethemselvesDelivery overcdn.optimizely.coman address the victims CSPapprovesTag, atob or evalthis is where the threepaths part
(Quelle: Optimizely status page entry, analyses by StepSecurity and Report URI)

Would it have worked on our domains?

Fetched again on 17 November, 11:58 to 12:01 UTC, with a browser identifier, checked against challenge markers: headers and HTML of our three domains. Optimizely references: none. integrity: null, crossorigin: null. The allow list is the only brake.

Short answer: a script-src with a host list stops the tag and the Base64 form, but neither the inline copy under 'unsafe-inline' nor the address swap in the page text.

Site (hosts per directive)script-srcconnect-src
senn-logistik.com14 hosts, 10 of them external, among them Stripe, PayPal, Leadinfo14 hosts, no return channel for stolen data
senn-tech.com2 hosts, neither of them externalself plus ip.senn-tech.com
senn-gruppe.comno script-src in the headerabsent

Unit: hosts per directive. Measured 17 November 2026, 11:58 to 12:01 UTC. Platform: plain HTTP fetch of the delivered headers and HTML with a browser identifier, no browser execution.

The finding concerns the policy; no browser ran these pages. Forms one and two would have failed at script-src on senn-logistik.com and senn-tech.com, because cdn.optimizely.com appears in neither list. On senn-gruppe.com the source restriction is missing altogether, and there all three branches would have run. The inline copy would also have run on the other two pages, since both carry 'unsafe-inline'. Our first reading of the eval branch was wrong: none of the three policies contains unsafe-eval, so on those policies that branch would have been refused. Report URI notes that 'strict-dynamic' and 'unsafe-eval' largely switch a policy off, and names Optimizely’s documentation as the source of the advice to allow unsafe-eval. The Optimizely CSP page, fetched on 19 November 2026 (state of 17 November), does not mention unsafe-eval; its example sets a nonce with 'strict-dynamic', and per StepSecurity that combination lets scripts added by a trusted script through. At HSBC the policy blocked the remote track.js and the data fetch according to Report URI, while the embedded copy ran anyway. Ten external hosts stand in script-src on senn-logistik.com while the homepage loads 9 scripts from our own zone: wider than the need, and that width was enough. No new approval was required, only an existing one.

Where a policy breaks the chainscript-src checks thetargetstops tag and atob formInline copy asfallbackruns with unsafe-inline orunsafe-evalconnect-src checksegressstops cards, cookies,screenshotsAddress swap in pagetextneeds no network, staysunchecked
(Quelle: Report URI, analysis of the policies of 256 hosts)

Checking a snippet by hand does not scale

We measured one public snippet address from that delivery, https://cdn.optimizely.com/js/6395285578973184.js, whose owner we do not know. The answer: 200, 312,216 bytes of minified code, text/javascript, Cloudflare with cf-cache-status: HIT, cache-control: max-age=120, access-control-allow-origin: *, plus x-amz-meta-revision: 29 and last-modified: Mon, 16 Nov 2026 21:38:09 GMT. The markers named in the public analyses (window._k7, data-channel-code, BIGNIAO) are absent, so we treat this snippet as clean. Nobody verifies that by eye: 312 kilobytes of code, one revision, 120 seconds of cache. What is needed is a hash per fetch across a timeline plus an alarm on every change of last-modified, as StepSecurity and Report URI also recommend.

Subresource Integrity pins a hash, and the snippet keeps changing

The integrity attribute on a <script> tag carries the hash of the expected content; if it does not match, the browser refuses to load the file and returns a network error. MDN names exactly the supply-chain attack as the use case, and the specification stands as a Working Draft of 20 March 2026. For the Optimizely snippet it does not work: StepSecurity and Mark Stott in the Optimizely community blog both state that the snippet changes with every published experiment and so cannot be pinned to one hash. Our integrity: null measurement had no consequence here, because none of the three sites loads an Optimizely script. The problem is not specific to Optimizely: the NDSS study Who's Hosting the Block Party? (Steffens et al., 2021; 10,000 top sites, 12 weeks) found that most sites include third-party code that requires 'unsafe-inline' (75 percent) or 'unsafe-eval' (61 percent), or makes 'strict-dynamic' impossible without breaking functionality (76 percent). On more than half of the sites SRI could not be used fully because third-party content changes. What remains is monitoring from outside: a hash per fetch and an alarm on every change.

The retreat visible in the registry

We queried eight domains from the published lists through rdap.org (17 November, 11:47 to 11:51 UTC), plus resolution and HTTP status through two resolvers (12:07 UTC). Two records confirm the takedown to the minute, the one point where our measurement agrees with an independent analysis: 1688360.cc, registered 16 November 05:23:08 UTC, now sits on client hold with its last change at 17 November 00:37:51 UTC, where Report URI notes 00:37 UTC. webshell.lol carries client hold from 03:37:29 UTC, where that analysis says 03:37 UTC. Both also answer NXDOMAIN through Google DNS, so the shutdown is real and independent of our network path.

DomainRegistration (UTC)StatusHTTP on 17 Nov, 12:07
eil786npto9u5bq.icu16 Nov 2026 16:38:42add periodno answer
linkforge.cc11 Apr 2026 22:36:28transfer prohibited403 Cloudflare block page
5vobauua3ojpklx.icu9 Nov 2026 17:52:42transfer prohibitedno answer, A record 34.96.213.201

ios-hack.icu and q1ls1ho9x7wsxlu.icu (16 November, 23:39 and 23:44 UTC, add period) are not in that table, because their HTTP status is a story of its own: both answer 200 on every path, including /track.js, with 5,771 bytes of text/html, titled "Sinkholed by Censys", no JavaScript, no marker. The infrastructure sits in a research sinkhole, and the trap belongs here: a 200 on a malicious path is not a hit; only content type and first bytes say what was delivered. By the afternoon of 17 November no name from our list was still active.

Four delivery routes, not one

Hudson Rock extended its report on 17 November: poisoned scripts were also served through Kissmetrics and StorifyMe (archived states from 13 November, three days before the Optimizely builds described here) and minute-ly (16 November, 16:25 UTC). Hudson Rock itself writes that this establishes neither four independent breaches nor a single operator. No vendor block list covers that; the check has to catch every third-party snippet we embed. The post on npm trusting a repository name and GitLab tripping over a regular expression asks the same trust question for package sources, and the security and privacy topic hub collects further cases.

Numbers with different denominators

Report URI names 1,611 snippets fetched, 188 of them infected, across 178 organisations. In short, per Report URI 24 of 159 reported hosts blocked the loader and 68 had no policy. It crawled the policies of 256 hosts of those organisations and reports 159 of them: 68 without any policy, 50 allowed both loader and exfiltration, 17 blocked the exfiltration while the loader still swapped the address in the text, and 24 blocked the loader. The 15 percent refers to those 159; measured against 256 it would be 9 percent. Where the other 97 hosts fall, the source does not say. Those 24 differ by a stricter script-src. Dividing 178 organisations by 256 hosts yields a protection rate never meant that way. The vendor’s initial report says sites with a strong policy were protected; the only number for that comes from this analysis.

Of 159 reported hosts, 24 blocked the loaderNo policy68 · 68 hostsLoader and egress allowed50 · 50 hostsOnly egress blocked17 · 17 hostsLoader blocked24 · 24 hosts070
The four groups add up to 159 of 256 crawled hosts; 24 of them are 15 percent. (source: Report URI, Scott Helme, 16 Nov 2026)

The money follows the same rule. Report URI counted roughly 7,040 dollars in incoming transfers at 19:25 UTC, of which 5,380 on the Bitcoin address: 0.08298 BTC from four transfers. Our own look on the morning of 17 November: 0.08528782 BTC received, closing balance zero. That fits the update at 09:17 UTC with 0.08528 BTC across five payments and 5,489 dollars, and the closing balance belongs to the onward transfers: 0.0852 BTC at 05:39 UTC to bc1q5fljq9c4g7gsjgg4etpn6k3lyusahcvjfv0u99, 17.02 LTC at 04:11 UTC, and 116 USDT at 04:13 UTC to a TRON address, in that same analysis in use since March and early November. OPTIMJACK is a research name, not a name from Optimizely.

What Optimizely Web customers should check now

ReversingLabs (17 November 2026) advises reviewing Project JavaScript and experiment code with their change history, including encoded and inline code, listing every snippet and the domain serving it (customer-hosted copies included), checking accounts, API tokens and integrations, and preserving the logs from 16 November: script responses, CSP reports, proxy and DNS logs. According to ReversingLabs the payload did not collect cardholder names or billing addresses and could not read cross-origin hosted payment iframes; most sites were exposed for about 16 hours. That is a security vendor's assessment, which we did not re-measure.

What we are not claiming

None of the named brands has reported data leaving its systems, and we measured none of those sites ourselves: the BBC, Zoom and HSBC attributions come from the analyses cited here. Hudson Rock sees a mobile branch toward Coruna and DarkSword in its samples but writes itself that the evidence does not establish a compromise of the backend systems it names.

What we changed here

First: a report-only twin of our three policies (header Content-Security-Policy-Report-Only, defined in CSP Level 3) with a report target, so we see violations of script-src or connect-src before changing anything. Second: we are reviewing which of the 10 external approvals on senn-logistik.com are needed and removing the rest. Third: the endpoint rule set gains the observation "browser process connects to an unknown address", because here the browser is the attacker process. Removing 'unsafe-inline' is the fourth item and the costliest, because our Next.js delivery emits inline blocks. It stays open. Anyone who needs such a policy with a reporting path for their own sites will find the frame under IT operations: virtualization, network, security and backup.

Further Reading

Questions?
Does a Content Security Policy protect against a supply-chain attack like the one at Optimizely?+

Only in part. A policy checks which address a script is loaded from, not what an already approved script does. The attacker wrote code into a script served under cdn.optimizely.com, an address the affected sites had approved themselves. A load attempt against a foreign address would have failed at the script-src field, an inline copy of the loader under 'unsafe-inline' would not. Of the 159 hosts Report URI reports, 24 blocked the loader.

Can a Subresource Integrity hash protect an Optimizely snippet?+

Practically not. The snippet changes with every published experiment, so a fixed hash in the integrity attribute would break loading at the next experiment. We fetched a snippet from cdn.optimizely.com: 312,216 bytes of minified code, Cloudflare with max-age=120, revision 29. At that size and cadence, a hash per fetch across a timeline with an alarm on every change helps; reading the file does not.

What happened in the Optimizely attack on 16 November 2026, and how many were affected?+

According to Optimizely, an attacker exploited a flaw in the Web Experimentation sign-up flow and wrote malicious code into the Project JavaScript of some customers. Report URI counted 188 infected snippets at 178 organisations among 1,611 snippets fetched. The status page named fewer than 5 percent of customers on 18 November 2026. The vishing incident Optimizely disclosed on 23 February 2026 is a different case.

Did the three Senn sites use Optimizely?+

No. We searched the delivered pages of senn-gruppe.com, senn-logistik.com and senn-tech.com on 17 November 2026: no reference to optimizely.com, no script from that address. That is an absence finding for the statically embedded scripts of these three pages, not a statement about other systems in the group.