# CVE-2026-94545 - .js ` /og` ImageResponse SVG Injection (GHSA-vcvr-r3jv-pc5j)
Security-research material for reproducing and validating **CVE-2026-94545**,
the .js `ImageResponse` SVG-serialization injection (upstream: Satori
**GHSA-wx4j-mvgx-mqwp**), in an isolated, loopback-only lab.
- **Affected:** .js `>=16.2.0 attacker
controls the embedded SVG document; RCE per the advisory depends on the
- **Upstream:** [vercel/satori GHSA-wx4j-mvgx-mqwp](
> **Use only on systems you own or are explicitly authorized to test.** The
> scanner is non-destructive: it renders an image and counts pixels. Nothing
> is written, stored, or executed on the target.
Against the advisory's exact route pattern (attacker value into
16.3.5 (vulnerable): injected 500x150 red rect rendered into the PNG
(75,000 #FF0000 pixels measured in the response)
16.3.6 (patched) : value escaped, 0 red pixels
satori 0.25.0 (vuln) : payload present raw in the serialized SVG string
satori 0.33.5 (fixed) : payload escaped to
This proves the injection primitive end-to-end (serializer -> embedded SVG ->
rasterized output). It does not demonstrate remote code execution: the
advisory ties RCE to the downstream SVG parser, whose vulnerable details are
not public at the time of writing. No file read, SSRF, or code execution is
## Root cause (verified against the fix commit)
Satori serializes `` nodes with `translateSVGNodeToSVGString()`
(`src/handler/preprocess.ts`), which interpolated into the output XML without
1. Text nodes returned via `String(node)` - unescaped (now `escapeXMLText`).
2. Attribute values interpolated raw inside quotes - `"` breaks out.
3. Attribute names interpolated raw - space/quote injects new attributes.
5. The embedded-SVG data-URL encoder did not encode `&`, so inner XML
6. `expand.ts` passed internal style properties (prefix `_`) into serialized
The fix ([satori@26a52aff](
routes everything through a hardening `buildXMLString()` that escapes values
and validates XML names. This lab verifies all of it behaviorally; the exact
request and pixel evidence are in `poc/`.
Requirements: Docker with Compose v2, Python 3.10+ on the host (for the
scanner), network access to npm on first run.
Builds the vulnerable app (vendored ` /og` from ` @16.3.5`), requires
the marker rect to render in the HTTP response, then the patched app
(` @16.3.6`) and requires its absence.
[*] #FF0000 pixels: 75000 (threshold 10000)
[!!!] INJECTION CONFIRMED: the attacker value escaped and
became a rendered SVG element in the generated image.
./lab start vulnerable # 16.3.5 app on
./lab test # assert injection renders
./lab test # assert absence
Run the scanner by hand (loopback default; authorized external targets need
python3 poc/scan.py
python3 poc/scan.py --param value \
--allow-authorized-non-loopback [--insecure]
## How the detection works (and why it is safe)
The payload is balanced XML carrying a 500x150 pure-red ``. On a
vulnerable endpoint the value escapes `` and the rect becomes a real
element in the embedded SVG document; the scanner counts exact `#FF0000`
pixels in the returned PNG (75,000 expected; threshold 10,000). On a patched
endpoint the value is escaped text and no red renders. The probe changes
nothing on the target: one GET request, one image render.
Note the carrier is deliberately a ``, not text: text inside the
embedded SVG has no loadable font in the renderer and would be invisible
The proof of this vulnerability is the pixel content of a rendered raster,
and nuclei matchers cannot decode PNGs. A size-differential template was
built and rejected: on a patched app the escaped payload still renders (as
literal inert text), so benign-vs-payload sizes differ on patched systems
too, and the separation between vulnerable and patched deltas is
app-specific compression noise. Shipping a matcher that fires on patched
systems is worse than shipping none. Use `poc/scan.py`; it evaluates the
| .js 16.3.5 (vendored @vercel/og) | Injection reproduced end-to-end over HTTP |
| .js 16.3.6 (vendored @vercel/og) | Patched negative control reproduced |
| satori 0.25.0 (npm) | Raw injection in serialized SVG (source level) |
| satori 0.33.5 (npm, 22 Sep 2026) | Escaped (source level) |
| .js 16.2.0-16.3.4 | Affected per advisory; not individually tested |
| .js 15.x | Not affected by the RCE issue per vendor blog |
│ └── server.mjs # advisory route pattern over loopback HTTP
└── scan.py # non-destructive scanner (Python stdlib: urllib + zlib)
Not committed: `assets/` (npm-packed tarballs + extracted bundles),
| SVG-serialization injection in satori < fix | Proven (source + behavior) |
| End-to-end rendering of injected markup via /og 16.3.5 | Proven |
| Patched negative control (16.3.6 / satori 0.33.5) | Proven |
| Attribute/style carriers (same root cause) | Root cause confirmed in fix diff; text carrier proven |
| SSRF / file read via injected resource references | Not tested |
| Remote code execution | Not claimed (depends on undisclosed downstream parser details) |
- [GHSA-vcvr-r3jv-pc5j ( .js)](
- [GHSA-wx4j-mvgx-mqwp (Satori)](
- [ .js security update, 22 Sep 2026](
- [Satori fix commit 26a52aff](
- [Netlify customer advisory (CVE-2026-94545)](
Run this repository only on systems you own or are explicitly authorized to
test. The lab driver refuses variants other than `vulnerable`/`patched`, the
scanner refuses non-loopback targets unless
`--allow-authorized-non-loopback` is supplied, and the Compose port is bound
See [SECURITY.md](SECURITY.md) for disclosure handling. This repository does
not include production targets, customer data, access tokens, or evidence
The full story
This article is one source in a clustered incident — the cluster page carries the summary, timeline and every other outlet covering it.
