CVE-2026-75650 (StyleSmuggler): unauthenticated RCE via template injection in Magento Open Source / Adobe Commerce 2.4.4–2.4.9 — b63 showcase, lab-verified with patched-build negative test
[b63] b63 -- Agentic Exploit Foundry
cve=CVE-2026-75650 class=unauthenticated remote code execution (template-directive chain, CWE-1336-style flows into include) verification=VERIFIED -- demonstrated on real Magento Open Source lab twins; patched twin refuses causally tested=2.4.8-p5 (this repo's lab) and 2.4.9 (cross-track lab build)
Unauthenticated remote code execution in Magento Open Source / Adobe Commerce:
a three-gap chain through the template/email subsystem ends with attacker PHP
included and executed in the web process, with command stdout returned inside
the HTTP response. Adobe assigned it VULN-39341 / APSB26-146; CISA lists it in
KEV. Full analysis: 05_deliver/writeup.md.
uid=33(www-data) with stdout and
app/etc/env.php material in the HTTP response, on a stock Magento Open
Source 2.4.8-p5 twin; the patched twin (same tree + Adobe VULN-39341 /
APSB26-146 patches, exactly one variable) completes the request chain but
refuses execution. Evidence: 04_verify/ (INDEX governs; superseded and
invalid captures are governed, not deleted).<?php exit; ?>, payload
neutralized to < ?=system(...); vulnerable: 6226-byte response with
uid=33(www-data). Reliability: 10/10 + 10/10 (vuln, opt/readable) vs 5/5 +
5/5 explicit refusals (patched). See
04_verify/11-11-reliability-postfix-10x10-negatives.txt and
04_verify/crosstrack-b/ (second independent lab, Magento 2.4.9).02_lab/lab/ lab recipe tree: compose + twin installers + Adobe patch set
03_exploit/exploit_opt/ minimal 5-request driver (stdlib Python 3)
03_exploit/exploit_readable/ same chain, named stages, commented path by path
04_verify/ captured evidence transcripts + INDEX.txt (governance)
04_verify/crosstrack-b/ A/B causality transcript from the second lab build
05_deliver/ writeup, usage (this repo's command reference), lab recipe
MANIFEST.txt sha256 over every shipped file (frozen after the scrub)
Prereqs: Docker + compose, Python 3 (stdlib only -- no pip installs). The lab
binds loopback only (127.0.0.1:18085 vulnerable / 18086 patched). Bring-up takes
~15-25 minutes for both twins to answer /graphql with 200.
# 1. build the twin lab (vulnerable + patched, from the Mage-OS mirror)
cd 02_lab/lab && docker compose -p cve-2026-75650-lab up -d
# readiness probe for both twins is in 05_deliver/lab-recipe.md
# 2. vulnerable twin -- RCE captured (exit 0, stdout in the response)
python3 03_exploit/exploit_opt/exploit.py --target http://127.0.0.1:18085
python3 03_exploit/exploit_opt/exploit.py --target http://127.0.0.1:18085 --cmd 'uname -a'
# 3. patched twin -- same request, causal refusal (exit 1, no execution receipt)
python3 03_exploit/exploit_opt/exploit.py --target http://127.0.0.1:18086; echo "exit=$?"
Exit-code contract (both drivers): 0 RCE captured | 1 chain completed, no
execution receipt (patched/non-vulnerable) | 2 usage | 3 transport failure.
Full command reference: 05_deliver/README.md;
lab bring-up/down and gate checks: 05_deliver/lab-recipe.md.
| State | Expectation |
|---|---|
| vulnerable 2.4.8-p5 / 2.4.9 twin | exit 0, uid=33(www-data) + command stdout inside the HTTP response |
| patched twin (Adobe VULN-39341 / APSB26-146, same tree) | exit 1, five fix gates present, stage-1 log poisoning still lands (guard is the only difference) |
setTemplateStyles vs unvalidated call site)
is documented from the vendor's patch text, not separately exercised as a
standalone unit; the end-to-end chain and the gate greps are captured
(04_verify/04-03b-gate-d-settemplate-styles-context.txt).04_verify/INDEX.txt so the correction trail stays
visible; only CURRENT rows carry claims.This repository is published for education and defence research only, to be used in an authorised laboratory environment on systems you own or are explicitly contracted to test. The exploit targets a lab twin on loopback. Running it against any third-party or production Magento instance without written authorisation is unlawful in most jurisdictions. The authors publish this as a lure-free technical record: no infrastructure, no payload delivery beyond the lab, no support for offence.
MIT -- see LICENSE.
MANIFEST.txt at the repository root lists sha256 over every shipped file as
frozen for publication (generated after the identity scrub; it supersedes the
run-time BUNDLE.sha256, which is not shipped).