Skip to content
Hudson Rock said

Hudson Rock said

www.infostealers.com • November 17, 2026

The confirmed compromise is malicious code inside responses served from third-party providers’ own script and widget endpoints. The affected files are not simply innocent names appearing in a request’s initiator stack. Kirk’s expanded evidence records the response bytes, size and SHA-256 for 15 website captures, spanning 16 vendor responses. Hudson Rock independently matched those hashes to the supplied archive records and directly recovered nine poisoned vendor files. [14]

In Optimizely’s case, the injection sits inside customer projectJS configurations, alongside legitimate experimentation code. Some versions use a plain loader, others encode the URL, and the BBC and HSBC UAE specimens embed a complete fallback payload. The code is not uniformly an identical line at the end of every customer file.

The shared responses and repeated injection establish a multi-customer delivery problem. They do not identify whether the original access involved provider infrastructure, publishing credentials, integrations or separate customer accounts. The evidence also does not establish a compromise of the named websites’ backend systems or every customer of these providers.

Hudson Rock also reviewed Amirul Cyber’s OptimJack research . Its asset leads helped expand the original sample set. Kirk’s later response-body evidence strengthens the findings for several sites that the first version of this report treated as researcher-reported leads. [1] [14]

Which websites are linked to poisoned responses?

The evidence below separates direct sample recovery from Kirk’s archived response-body verification . Both are more specific than a domain appearing somewhere in a dependency chain. Neither proves that every visitor loaded the malicious version or that a visitor lost data.

Hudson Rock’s original Optimizely sample set

Hootsuite and Talkwalker one poisoned project. Talkwalker’s homepage HTML directly referenced 4894383279964160.js . This makes nine linked web properties in the table, while the number of distinct recovered Optimizely assets remains eight.

Additional sites in Kirk’s response-body verification

Kirk’s 15-capture table overlaps our findings for BBC, Zoom and GitLab and adds the 12 properties below. Together, the two tables identify 21 named web properties with the stated evidence. This is a documented subset of the activity, not an incident-wide victim count. [14]

CNN, Salesforce, Mutual of Omaha and Fox News now have specific historical Optimizely response evidence. In particular, the Fox News file is /js/5728063363743744.js . The result does not depend on interpreting a Confiant wrapper or Salesforce helper as the compromised vendor.

Kirk reports 51 traced sites in the wider investigation, but his supplied body verification covers the 15 captures above. Broader initiator-only leads are not silently into this verified subset. The HSBC UAE asset recovered by Hudson Rock is also distinct from the HSBC Singapore insurance project mentioned in wider reporting.

Historical evidence remains relevant when a later response changes. Hudson Rock’s later copies of several reported assets lacked the known injection; those later copies do not invalidate the earlier archive hashes and body findings.

Large audiences, specific browser-level risks

The affected asset associations span news, banking, communications, software and business services. Their importance comes from the audiences that trust these websites and the actions a browser may perform after it loads the script.

These figures must not be added together or presented as victims. Monthly audience, registered accounts, global banking customers, network connections and enterprise customers are different measures. None measures the number of people who encountered this payload.

The wallet and card modules operate through a page’s accessible content and APIs, subject to same-origin and HttpOnly restrictions. Separately, a functioning browser or device exploit could extend access beyond that page. The iOS delivery evidence below raises that additional risk but does not demonstrate successful exploitation or establish which data a later stage obtained. [11] [12]

The mobile branch connects to Coruna and DarkSword infrastructure

The original recovered track.js already showed iOS-specific secondary loaders. Kirk’s additional work follows that branch further: into a router associated with Coruna and DarkSword delivery . Hudson Rock independently reviewed the two supplied scan records and their response hashes. [14] [19] [20]

What the two iOS-directed captures actually show

The 21:08:43 UTC GitLab scan presented an iOS 18.6 Safari user-agent string. Its records show successful responses for both boot.js hosts, 5vobauua3ojpklx[.]icu/D11/b.html , and eil786npto9u5bq[.]icu/JIUJIU888/index.js . Other requests, including D12/b.html and an ios-hack[.]icu IP-sync call, appear without a completed response. [19]

The 21:12:49 UTC scan presented an iOS 17.5 Safari user-agent string and received eil786npto9u5bq[.]icu/A8/a.html and A8/34ef644ff217a50beadde64dc8e23a7d.js . [20]

These are delivery observations, not demonstrated iPhone compromises. URLScan uses headless Google Chrome and allows a custom user-agent string. Changing that string does not create Safari’s WebKit engine or an iOS kernel. Successful HTTP responses therefore establish what the router served, not whether an exploit succeeded on an Apple device. [16]

Why the Coruna and DarkSword connection is credible

Kirk’s notes identify ds_ visitor keys stored as ds_vk , configuration object __DAILI_CFG , and a hidden 320 by 568 frame named chainFrame with data-daili-gate="v1" . He reports XOR-encoded API and C2 host strings in the router, decoded with key d7K9mP2x . The entry code collects device and landing-page telemetry and uses IP-sync and execution-state endpoints. His D11 analysis also describes a once-per-24-hour frame gate and a 30-minute pause after two failures. These are researcher-supplied static findings, not timings measured on victim devices. [14]

The same channel and infrastructure support linked delivery activity across the November 13 and November 16 observations . They do not prove that one person controls all affected vendors, identify the kit’s buyer, or establish a shared initial-access mechanism. A commercial kit can be reused by multiple operators.

C10 does not establish a working iOS 26 exploit

In its November 5 report, updated November 6, Ohey documented a related router with C10 disabled and unavailable on the hosts checked. Earlier staged C10 files contained generated offsets and placeholder values. Ohey also found the combined kit advertised as Caishen iOS C2 V2. Those historical observations do not determine the current state of different campaign hosts. [15]

Kirk’s current notes contain a version-26 C10 branch, but the supplied iOS 26.3 test did not reach track.js . A configured route is not proof of deployed, functional exploit code, and the failed delivery test is not proof that the device version is safe.

There is also a browser-identification limit: Safari 26 freezes the operating-system token in its user-agent string to an older value. A router’s version labels cannot be equated with real-device behavior without inspecting its parser and validating the relevant environment. [17] [18]

The supported conclusion is a connection to exploit-kit delivery infrastructure. The current evidence does not establish a zero-day, successful device exploitation, or collection of specific device data by unseen downstream stages.

Blocking the attacker domain can leave the malware running

The BBC and HSBC UAE scripts contain more than a reference to an external file. Each embeds the same 24,392-byte malicious payload in Base64 , already present inside the Optimizely response.

The loader decodes that payload locally, then attempts to fetch remote track.js . If the request errors, or has not completed after 2.5 seconds , it attempts to execute the embedded copy. A later check can retry retrieval through fetch and XMLHttpRequest when the expected execution marker is absent.

A DNS or network block is useful, but it does not remove the poisoned trusted script. It may stop requests to the blocked host while leaving local wallet-address substitution available from the embedded code. The separate iOS hosts are additional destinations. A site’s CSP may constrain execution paths, but containment should verify removal of the injected project code and the bytes delivered to visitors.

The live and embedded payloads are different variants. The live sample has additional request and function hooks, raw-format TON handling and a second iOS loading path. Their card-skimming module is byte-identical. The comparison establishes different capabilities, not a proven development chronology.

Wallet substitution and card skimming accompany the mobile branch

The behavior below comes from Hudson Rock’s recovered live payload and decoded fallback. Their wallet and card modules are not restricted to desktop visitors; iOS detection adds secondary loading rather than replacing those modules.

1. Cryptocurrency address replacement

The script contains hardcoded replacement destinations for Ethereum/EVM-style addresses, Bitcoin, TRON, TON, Solana, Litecoin, Bitcoin Cash and Dogecoin. It recognizes address formats and validates checksums for several families before substituting the destination. EVM recognition is format-based, while Solana matching includes heuristics.

Its targets include visible text, input and textarea values, selected element attributes, links, copy events and clipboard API calls. The live variant also attempts to modify address-bearing strings in some request bodies, request URLs, /history operations and sharing calls. Mutation observation and a 50 ms scan interval while the page is visible help it revisit content that changes after load.

The effect can be a payment sent to a destination the attacker placed into the page or copied value. The code does not need to obtain a private key to attempt this diversion. In the recovered samples, we did not identify seed extraction, wallet-signing calls or a mechanism proving an automatic on-chain transfer.

Address replacement is not uniformly restricted to payment-related keywords. In the core text-replacement function, context affects whether an original address is recorded for collection; valid matches can still be replaced without that logging condition. URL processing has its own additional checks.

2. Payment-card skimming

The card module hooks form submissions and POST bodies sent through fetch and XMLHttpRequest. It can inspect JSON-like bodies, URL-encoded parameters and FormData for a card number, expiry and security code.

Before attempting transmission, it requires a recognized card-number prefix, a Luhn-valid number, an expiry matching its checks and a three- or four-digit CVV/CVC. It then attempts to send the complete tuple to 1688360[.]cc/card.php , using sendBeacon when available.

This is implemented card-skimming behavior, not evidence that every affected page presented a payment form or that a card was actually transmitted. The code does not demonstrate bypass of cross-origin hosted-payment-frame isolation. [12]

3. Conditional cookie and page-image collection

A separate routine prepares a POST to 1688360[.]cc/hit.php containing the page URL, JavaScript-readable cookies, recorded address matches and an optional JPEG rendering of the viewport. It uses html2canvas , requesting the helper from the attacker host if a usable instance is not already present.

This path is conditional, not a beacon that runs for every visitor. The screenshot routine requires recorded address matches, follows relevant visible changes and is guarded to run once per page. If rendering fails, the routine can attempt to send its other fields without an image.

The constructed fields include up to 500 characters of URL and 8,000 characters from document.cookie . A readable session cookie could therefore be included, but HttpOnly cookie values remain unavailable to this JavaScript API . We did not identify a localStorage extraction routine. [11]

The image is a rendering of page content. It is not an operating-system screenshot and does not by itself show access to unrelated tabs or applications.

4. iOS-specific loading in the recovered variant

The live sample recognizes iPhone, iPad and iPod user agents and an iPad-style Macintosh/touch combination. It inserts two boot.js references with the 02111110 channel marker. A separate branch adds JIUJIU888/weifile.html when the parsed major/minor iOS version is 18.6 or below; patch versions are not distinguished by that check.

Hudson Rock’s initial direct requests received small Site Unavailable pages. Kirk’s later supplied captures show that other request conditions obtained router and entry-stage responses. An unavailable response in one retrieval is therefore not evidence that the downstream delivery was universally inactive.

No mining implementation was identified in the recovered variants. The SHA-256 code is used for cryptocurrency-address checksum validation. Its presence does not establish cryptocurrency mining.

The evidence spans multiple dates and payload versions

The BBC observations illustrate why response versions and timestamps matter. A malicious file received at 21:00 and a different response received by another researcher at 21:10 can reflect an intervening change or different cache responses. Neither observation establishes the state of every edge cache, page or already-open browser session.

The early GitLab scan records delivery without a card.php or hit.php request. The later iOS-directed scans provide additional routing evidence. Delivery, completed response, exploit execution and data theft are separate findings and must not be treated as interchangeable.

Domains, paths, wallets, hashes and detection pivots

Evidence status is part of each indicator. An endpoint in code, an attempted request, a successful response and verified data collection are different observations. Related historical Ohey infrastructure is not automatically into the current campaign’s blocklist.

ipv6.icanhazip.com appears in Kirk’s description as a public IP-discovery service. Its use by the router does not make that legitimate service attacker-owned or a reliable standalone indicator.

Hardcoded cryptocurrency destinations

Hashes of directly recovered malicious specimens

Mobile-router and stage hashes from archive records

Kirk reports 12,088 bytes for the first boot script and 12,637 for the second. The two hashes above are bound to their exact URLs in the scan records. They are distinct from the separately reported boot sample in the historical-hash table below. [14] [19] [20]

Earlier and externally reported payload builds

Additional reported-only domains

These four domains remain alerts supplied through the original researcher posts, without a matching request or body in the records examined here. l7bcsnwc7nzr3uo[.]icu has moved into the recorded-request table above.

Code, storage and routing markers

Use these strings in combination with origin, path, content and timing. They are not individually reliable signatures. The card-deduplication keys can contain card details and require appropriate evidence handling. The C10 marker is a hunting lead, not proof that its route delivered a working exploit.

Remove the delivery code and investigate both forms of exposure

Preserve and remove the affected third-party code. Retain actual response bytes, hashes, headers, timestamps and project revisions before restoring or disabling the relevant project or resource. Inspect Optimizely projectJS , the affected Kissmetrics/StorifyMe endpoints and minute-ly’s loader where used. Verify what visitors actually receive after changes.

Block both delivery paths and the observed downstream hosts. Cover 1688360[.]cc , webshell[.]lol and the current mobile-router indicators. Domain blocks are only one layer: the BBC and HSBC UAE specimens already contain a complete embedded payload.

Preserve responses and network traces, not just the final page DOM. The minute-ly specimen schedules removal of its own and related script elements after three seconds. Removing elements does not undo executed code, but it can make a later DOM-only inspection incomplete.

Separate financial-theft exposure from mobile-stage delivery. Hunt for card.php and hit.php POST attempts, boot/entry-stage requests and state/telemetry calls as separate events. An endpoint named pe_ok or a result parameter is not independent proof that privilege escalation occurred.

Scope sensitive pages and device exposure individually. Determine whether payment, login, account or other sensitive paths loaded the affected file during the relevant period. Check applicable browser/device versions and seek device-level forensic evidence if exploit exposure is suspected. Refresh affected software and follow verified vendor guidance. Invalidate exposed sessions or coordinate payment-data response where the evidence supports it.

Investigate the publishing path. Review project changes, publication history, account activity, API credentials and integrations with the relevant providers. Shared malicious output establishes the integrity problem; it does not by itself reveal how the attacker gained access.

For visitors, the relevant questions are which page was loaded, when, on which device, and what was entered or copied. Cryptocurrency payment destinations should be checked through an independent trusted view. An organization-specific exposure notice should guide any payment or session response; membership in a large customer population alone does not establish compromise.

The new evidence changes the scope of the investigation. This activity combines ordinary browser theft with delivery infrastructure associated with iPhone exploit kits. Containment must address the poisoned trusted scripts and both downstream delivery paths, while incident reporting keeps requested stages, executed code and confirmed losses distinct.

What the combined evidence establishes

Hudson Rock retrieved JavaScript and supporting HTML as data, preserved bytes and SHA-256 hashes, decoded the embedded Base64 payload without evaluating it, compared variants and reviewed existing scan records. No malicious code was executed in Hudson Rock’s analysis, no new scan was submitted, and no collection endpoint was supplied victim data.

The expanded source is Kirk’s updated November 16 research note, supplied directly to Hudson Rock. It records 15 website captures containing 16 poisoned vendor responses. Hudson Rock independently corroborated all 16 recorded hashes and vendor URLs, and directly recovered the minute-ly loader in addition to the original eight Optimizely assets. For historical bodies unavailable in our retrieval, their code interpretation remains explicitly attributed to Kirk.

The two iOS-directed scan records also corroborate response hashes and route selection. URLScan’s Chrome engine makes these delivery tests, not real-device exploit validations. Kirk had not fully read the later DarkSword stages or decoded the Coruna module maps in the supplied account. The initial compromise mechanism, operator identity, visitor count and successful theft or device-compromise totals remain unresolved.

The GitHub issue referenced in the supplied notes was unavailable during this review and was not treated as independently reviewed public evidence. Audience metrics describe the organizations’ scale and cannot be added together or converted into incident counts.

Amirul Cyber: OptimJack research and reported page-to-script associations

urlscan: GitLab marketing-site capture, November 16, 2026, 16:17 UTC

urlscan: archived track.js response record, November 16, 2026, 19:03 UTC

Who Gets My Data: HSBC UAE browser-request evidence

BBC Studios: BBC.com audience scale in its 2026 Africa Enterprise announcement

GitLab: Q2 FY2027 results, published September 1, 2026

HSBC: company history and global customer count, updated November 11, 2026

Virgin Media O2: Q3 2026 operating metrics

Hootsuite: official company page, reviewed November 16, 2026

Cloudflare Radar: independent response hash and infrastructure record

MDN: document.cookie and HttpOnly limitations

MDN: same-origin policy and cross-origin frame access

urlscan: earlier track.js response record, November 16, 2026, 16:07 UTC

Kirk / Ohey: updated research notes supplied to Hudson Rock, November 16, 2026. Includes 15 archived response-body checks and iOS-directed scan findings. Source document SHA-256: 3c693a9e3117d84ba62a20012ba7d5c41f91ece27a5ec1b367e34bf7559fabdb

Ohey: DarkSword router’s iOS 26 route, C10, is switched off; published November 5, updated November 6, 2026

urlscan documentation: headless Chrome, custom user-agent support and point-in-time scan limitations

WebKit: Safari 26.0 user-agent changes

WebKit issue 298673: frozen OS token updated to 18.7

Kirk-supplied GitLab capture using an iOS 18.6 Safari user-agent string, November 16, 21:08:43 UTC

Kirk-supplied GitLab capture using an iOS 17.5 Safari user-agent string, November 16, 21:12:49 UTC