On 27 October 2026 CISA added CVE-2026-48356 to its Known Exploited Vulnerabilities catalog: an unrestricted file upload in Adobe Commerce and Magento Open Source. The security firm Sansec, which found it and named it PolyShell, had published it on 17 March. Mass scanning started two days later. Adobe shipped a fix for production release lines on 14 July. The KEV listing came 224 days after the public warning.

By the time of the listing, the bug was not the interesting part of the story any more. Sansec's 24 October report describes an operator re-infecting stores in October with no vulnerability at all, because 60 of the 83 stores on its newest domain were already compromised. For many stores the question stopped being "are we patched" months ago. It is "who is already in".

[INFO]
Primary sources: Sansec's PolyShell advisory (17 March 2026, with dated updates), its 30 March mass-attack report and its 24 October deepveindrilling.com report; Adobe's CVE record for CVE-2026-48356 (published 14 July, updated 28 October); CISA's KEV alert of 27 October. Adobe's bulletin APSB26-73 itself could not be opened (HTTP 403); affected and fixed versions here come from Adobe's CVE record, which links to it.
◈ interactive artifact
Patched Is Not Clean
A store backend to search: version, media folder, CMS blocks, admin users and server config. Flag Sansec's four indicators as you find them and watch the triage board answer the three questions patching does not.

//What the Bug Is

Magento lets a shopper attach a file to a cart item when a product has a file-type custom option - an engraving logo, a print file. Sansec found that the REST API's guest-cart routes (POST /V1/guest-carts/:cartId/items and PUT /V1/guest-carts/:cartId/items/:itemId) would accept such a file for any product. Three checks were missing: the option ID was never matched against the product's options, the upload ran even if the product had no file option, and there was no extension restriction. The only test was whether the file looked like an image, which a file that starts with an image header and continues as PHP passes. Hence the name: a polyglot shell.

Guest carts need no account, so neither does the upload. Files land in pub/media/custom_options/quote/. Sansec says the vulnerable code dates back to the first Magento 2 release. GraphQL uses a different code path and is not affected.

Whether an uploaded file becomes code execution depends on the web server. Per Sansec: with Adobe's sample nginx and Apache configurations on current 2.4.x lines, the uploaded file is not served or executed. It does execute on non-stock nginx that passes every .php request to FastCGI, and on Apache without php_flag engine 0 for the media directory. On older lines it can also be stored cross-site scripting. Even where it does not execute, the file stays on disk, waiting for a configuration change or a server migration.

//The Timeline

16 March 2026, 12:00 UTC - first probing, per Sansec, the day before its public advisory.
17 March - Sansec publishes PolyShell. Adobe had fixed it in the 2.4.9 pre-release branch only; there was no patch for production versions.
19 March - first attacks in the wild; automated mass scanning begins.
24 March - Sansec reports malicious PHP uploaded to 56.7% of the stores it sees.
30 March - a single wave injects a JavaScript loader into 471 stores in one hour, after uploading webshells through PolyShell. None were Sansec customers.
14 April - 82% of stores hit with malicious uploads, per Sansec.
12 May - Magento 2.4.9 released. Sansec says it carried the fix, with no backport to 2.4.8 or earlier.
14 July - Adobe publishes APSB26-73 and the CVE, with fixed -2026-jul releases for every supported line.
6 to 23 October - the same operator registers a new loader domain, puts it on 83 stores, then rewrites the injected blocks with a fingerprinting version.
24 October - Sansec publishes the deepveindrilling.com report.
27 October - CISA adds CVE-2026-48356 to KEV.

One caution on the percentages. Sansec's figures come from its own monitoring, and its advisory says "all stores" in one place and "all protected Magento stores" in another. They describe the stores Sansec can see, not every Magento store on the internet. They are still the best public measure of how fast an unpatched upload bug gets used.

//Two Descriptions of One Bug

Adobe's CVE record scores CVE-2026-48356 at 9.3 under CVSS 3.1 (AV:N/AC:L/PR:N/UI:R/S:C/C:H/I:H/A:N). Its description says exploitation "requires user interaction in that a victim must visit a maliciously crafted URL or interact with a compromised web page". That reads like the stored cross-site scripting outcome. Sansec's write-up describes an unauthenticated upload over the REST API, with code execution on some server configurations, and its attack data shows exactly that being done at scale with no victim clicking anything.

Both can be true at once - the score covers the worst case Adobe chose to model. But a defender reading only the CVE text would rank this lower than the attackers did. The record also lists plain 2.4.9 as affected, with 2.4.9-2026-jul as the fix, while Sansec says 2.4.9 already carried the fix in May. If you run 2.4.9, install the July release rather than rely on either reading. CISA's own enrichment on the record now marks exploitation as "active".

//Affected and Fixed

From Adobe's CVE record. Affected: Adobe Commerce 2.4.9, 2.4.8-p5, 2.4.7-p10, 2.4.6-p15, 2.4.5-p17, 2.4.4-p18 and earlier; Magento Open Source 2.4.9, 2.4.8-p5, 2.4.7-p10, 2.4.6-p15 and earlier; Adobe Commerce B2B 1.5.3, 1.5.2-p5, 1.4.2-p10, 1.3.4-p17, 1.3.3-p18 and earlier; Adobe Commerce Events 1.20.0 and earlier.

Fixed: 2.4.9-2026-jul, 2.4.8-2026-jul, 2.4.7-2026-jul, 2.4.6-2026-jul, and for Commerce also 2.4.5-2026-jul and 2.4.4-2026-jul. B2B: the matching -2026-jul releases. Events: 1.21.0. Lines older than those have no fix.

//Why Patching Is Not the End

Sansec has tracked one operator since July 2025 through five loader domains and 953 stores, mostly Magento. Its 30 March wave used PolyShell. Its October wave used nothing new: of 83 stores on the new domain, 60 were already infected by earlier loaders, and the operator simply rewrote the blocks it had planted. One store on 2.4.7-p10 received the loader an hour after a clean scan; Sansec notes it still lacked APSB26-73.

The loader lives in the database, in Magento's cms_block table, which Magento renders into storefront pages. The earlier version stored its payload in the visitor's browser localStorage under a key called fpCode and ran it on later page views without fetching anything, so blocking the domain does not clean browsers that already have it. The 23 October version sends a browser fingerprint first and lets the server decide what to serve, filtering out automated scanners. Sansec's own summary of the cleanup problem is blunt: "Deleting the script does not remove that access."

So the order matters. Patch to stop new uploads. Then look for what earlier uploads left behind, remove it, and rotate every credential the attacker could have read.

//What to Check

From Sansec's two reports:

Upload directory. find pub/media/custom_options/ -type f ! -name '.htaccess'. Any .php, .phtml or .phar there is a strong indicator of compromise; the October report also lists pub/media/customer_address/ and webshells disguised as PNG files.
Backdoor. find /var/www -name 'accesson.php' -type f. In March it was copied into assets/images/ under many top-level directories of a single store, so search the whole tree.
Database. Search cms_block and cms_page content for fpCode, se_referrer and localStorage.getItem, and check update_time for writes you did not make. Include table prefixes if you use them.
Accounts. Review admin_user for accounts nobody recognises, and look for unfamiliar processes.
After cleanup. Change database, admin and API integration passwords.

If you cannot patch today, Sansec's interim steps are to block the vulnerable API call at a WAF and to deny web access to pub/media/custom_options/ - an nginx location placed before any .php block, or an .htaccess with php_flag engine 0 and Deny from all on Apache. Denying the directory stops an uploaded file from being served. It does not stop the upload.

[IOC]
CVE: CVE-2026-48356 (PolyShell), CWE-434, CVSS 3.1 9.3 per Adobe; fixed by APSB26-73 (14 July 2026)
KEV: added 27 October 2026
Routes: POST /V1/guest-carts/:cartId/items, PUT /V1/guest-carts/:cartId/items/:itemId (unauthenticated)
Upload path: pub/media/custom_options/quote/
Backdoor: accesson.php
Loader domains (Sansec): deepveindrilling[.]com, additionalver[.]com, lanhd6549tdhse[.]top, neswnsssn[.]click, manulfie[.]click, hdfdsi84j3h[.]top, c85grt98sd[.]com
Payload hosts (Sansec): leqdger[.]click, crypto-insider[.]top
Also named in the March advisory: jslibrary[.]net, canevaslab[.]com
Browser artefact: localStorage key fpCode; database: cms_block.content

//Sourcing and Uncertainty

The mechanism, routes, server-configuration outcomes, attack percentages, March timeline, mitigations and detection commands are from Sansec's PolyShell advisory and its 30 March report. The campaign since July 2025, the 953 and 83 store counts, the 60 re-infected stores, the 2.4.7-p10 example, the loader behaviour and the cleanup checks are from Sansec's 24 October report. Affected and fixed versions, the CVSS vector, the user-interaction wording and CISA's "active" exploitation marking are from the CVE record (as fetched on 29 October). The KEV date and name are from CISA.

Where sources disagree. Adobe's CVE text frames this as a user-interaction bug; Sansec's data shows unauthenticated mass exploitation. Adobe lists 2.4.9 as affected; Sansec says 2.4.9 shipped the fix. Several secondary reports say Adobe's bulletin still states it is not aware of exploitation; that bulletin could not be opened here, so this article does not repeat the claim.

What is unknown. How many stores were compromised in total - Sansec's counts are of the stores it monitors and its per-domain figures overlap, so they cannot be added up. Who the operator is. Sansec does not attribute the campaign, and neither does this article. What the loader ultimately serves to visitors who pass its fingerprint check is not described in detail in the reports reviewed here.