CVE-2026-44589

Description

## Summary The `isBlockedUrl()` denylist introduced in `nuxt-og-image@6.2.5` to remediate **GHSA-pqhr-mp3f-hrpp** (Dmitry Prokhorov / Positive Technologies, March 2026) is incomplete. The patch advisory states "Decimal/hexadecimal IP encoding bypasses are also handled" — that part is true (Node's WHATWG URL parser canonicalizes those forms before validation), but the v6.2.5 implementation misses two independent surfaces in the latest release `6.4.8`: 1. **IPv6 prefix list is incomplete.** The IPv6 branch checks only `bare === "::1" || startsWith("fc") || startsWith("fd") || startsWith("fe80")`. It misses: - `[::ffff:7f00:1]` — IPv6-mapped IPv4 loopback in pure-hex form (RE_MAPPED_V4 regex requires dotted-quad). **Reaches 127.0.0.1 on a single-stack-IPv4 host with no other primitive needed.** - `[fec0::/10]` (RFC 3879 site-local — deprecated but still routable on legacy networks) - `[5f00::/16]` (RFC 9602 SRv6 SIDs) - `[3fff::/20]` (RFC 9637 IPv6 documentation v2) - `[64:ff9b:1::/48]` (RFC 8215 NAT64 local-use, including embedded IPv4 loopback `[64:ff9b:1::7f00:1]`) 2. **No redirect re-validation.** `isBlockedUrl` runs once on the initial `<img src>`. The subsequent `$fetch(decodedSrc, ...)` (ofetch, default redirect-follow) follows 30x responses with no second-pass validation. Any allowed origin that returns a 302 to an internal IP — S3 redirect rules, GCS, Azure, CloudFront, any user-content CDN where the attacker can place a single redirect — completes the SSRF. The net result is that the v6.2.5 SSRF advisory is bypassable in two distinct ways. The same root family as #29 / #38 (ipx) but in a **different code path with different gaps** — `nuxt-og-image` does not delegate to `ipx`, it ships its own validator, and that validator has fresh issues that survived the prior fix. ## Affected | Package | Version | Role | |------------------|-------------------|-----------------------------------------------------| | `nuxt-og-image` | `6.4.8` (latest) | default OG-image generator for Nuxt apps | | `@nuxtjs/og-image` (alias) | same | re-export, same code path | The vulnerable code lives in `dist/runtime/server/og-image/core/plugins/imageSrc.js` and is enforced for every `<img src>` (and `style="background-image: url(...)"`) inside an OG image component, on production builds (`!import.meta.dev`). ## Vulnerable code (`imageSrc.js`, verbatim) ```js function isPrivateIPv4(a, b) { if (a === 127) return true; if (a === 10) return true; if (a === 172 && b >= 16 && b <= 31) return true; if (a === 192 && b === 168) return true; if (a === 169 && b === 254) return true; if (a === 0) return true; return false; } function isBlockedUrl(url) { let parsed; try { parsed = new URL(url); } catch { return true; } if (parsed.protocol !== "http:" && parsed.protocol !== "https:") return true; const hostname = parsed.hostname.toLowerCase(); const bare = hostname.replace(RE_IPV6_BRACKETS, ""); if (bare === "localhost" || bare.endsWith(".localhost")) return true; const mappedV4 = bare.match(RE_MAPPED_V4); // /^::ffff:(\d+\.\d+\.\d+\.\d+)$/ const ip = mappedV4 ? mappedV4[1] : bare; const parts = ip.split("."); if (parts.length === 4 && parts.every((p) => RE_DIGIT_ONLY.test(p))) { /* dotted-decimal IPv4 path */ } if (RE_INT_IP.test(ip)) { /* single-integer IPv4 path */ } if (bare === "::1" || bare.startsWith("fc") || bare.startsWith("fd") || bare.startsWith("fe80")) return true; // ← gap: only 4 IPv6 prefixes return false; // ← everything else is "public" } // Then: async function doResolveSrcToBuffer(src, kind, ctx) { ... if (!import.meta.dev && isBlockedUrl(decodedSrc)) { return { blocked: true }; } const buffer = await $fetch(decodedSrc, { // ← follows 30x by default responseType: "arrayBuffer", timeout: fetchTimeout, }); ... } ``` Two distinct issues: - **The IPv6 prefix list is hand-rolled** (`fc`, `fd`, `fe80`, `::1`) and inherits no taxonomy from `ipaddr.js` or any RFC table. - **`$fetch` is `ofetch`**, which wraps Node `fetch()` with default `redirect: "follow"`. The validator does not run on the redirect target. ## Reproducer (verbatim, no host privilege) End-to-end test of `isBlockedUrl` on a corpus of internal-IP forms, paired with empirical `fetch()` confirming which forms actually reach loopback. Verbatim output: ``` isBlockedUrl? fetch reaches loopback? url ------------- ----------------------- --- ✓ blocked YES http://127.0.0.1:8765/ (control: dotted-decimal loopback) ✓ blocked YES http://localhost:8765/ (control) ✓ blocked no(ECONNREFUSED) http://[::1]:8765/ (control: IPv6 loopback) ✓ blocked no(EHOSTUNREACH) http://169.254.169.254:8765/ (control: AWS IMDS) ✓ blocked YES http://2130706433:8765/ (control: decimal-int IPv4) ✓ blocked YES http://0x7f000001:8765/ (control: hex-int IPv4) ✓ blocked YES http://0177.0.0.1:8765/ (control: octal — URL parser canonicalizes) ✓ blocked YES http://127.1:8765/ (control: shorthand — URL parser canonicalizes) ✗ NOT blocked YES http://[::ffff:7f00:1]:8765/ (BYPASS: IPv6-mapped, hex form) ✗ NOT blocked no(unreachable) http://[fec0::1]:8765/ (BYPASS: RFC 3879 site-local) ✗ NOT blocked no(unreachable) http://[5f00::1]:8765/ (BYPASS: RFC 9602 SRv6) ✗ NOT blocked no(unreachable) http://[3fff::1]:8765/ (BYPASS: RFC 9637 docs) ✗ NOT blocked no(unreachable) http://[64:ff9b:1::1]:8765/ (BYPASS: RFC 8215 NAT64) ✗ NOT blocked no(unreachable) http://[64:ff9b:1::7f00:1]:8765/ (BYPASS: NAT64 + embedded loopback) ``` The first six bypass rows say "✗ NOT blocked" — that is `isBlockedUrl` returning `false` (i.e., "this URL is fine to fetch") for each of those addresses. The "fetch reaches loopback" column shows that `[::ffff:7f00:1]` actually round-trips to 127.0.0.1 on a single-stack-IPv4 dev box; the four cluster ranges are unreachable on the dev box but succeed on dual-stack / k8s / NAT64 / SRv6 networks where any of these prefixes is internally bound. The "control" rows confirm the bypass set is minimal — the validator catches the obvious cases. The bypasses are the cases the prefix list forgot. ### Class 2: redirect amplifier `$fetch(url, { responseType: "arrayBuffer", timeout })` follows 30x by default. Confirmed empirically — `ofetch('http://lab.menna.website/test/redirect-to-loopback')` (where `lab.menna.website` returns `302 Location: http://127.0.0.1/`) ends with `<no response> fetch failed` after the connect attempt to `127.0.0.1:80`, proving the redirect was followed. On a target where the redirect destination has a service bound, the bytes round-trip back through the OG renderer. Same primitive as #29 / #38 (ipx redirect bypass), in a different validator. The fix recommendations for #29 also apply here, with the same trade-offs. ## Impact A Nuxt application that uses `nuxt-og-image` (the official-recommended OG generator) and includes any user-influenced URL in an OG component is vulnerable to SSRF that returns the bytes of the internal response as part of the rendered OG image: - **Class 1 directly:** `<img src="http://[::ffff:7f00:1]:PORT/path">` reaches 127.0.0.1 on the OG worker. If the dev's deployment has anything bound to loopback (admin dashboards, internal HTTP-RPC, Redis HTTP UI, anything running alongside the function on the same machine in self-hosted setups), it leaks. - **Class 1 cluster:** the IPv6 cluster ranges trigger only on dual-stack / k8s / NAT64 networks — but those are exactly the production targets where SSRF matters most. - **Class 2 redirect:** any allowed CDN with a redirect rule extends the reach to all RFC 1918 / loopback / link-local space. `nuxt-og-image` is the OG-image module recommended in Nuxt's official documentation; it is shipped with Nuxt UI templates and is one of the top-2 Nuxt modules by GitHub stars. The user-facing primitive in real apps is "title/avatar comes from a request param" — exactly the same `<NuxtLink to="/og?avatar=...">` pattern Nuxt docs encourage. ## Suggested fix Three non-exclusive options: 1. **Replace the hand-rolled IPv6 prefix list with `ipaddr.js`'s `range()` predicate** (or equivalent), then either: - explicitly deny the four cluster ranges that `ipaddr.js` currently misses (`fec0::/10`, `5f00::/16`, `3fff::/20`, `64:ff9b:1::/48`), or - wait for the `ipaddr.js` upstream patch (see Vercel #27 — same gap, separately disclosed) and bump. - In any case, also catch `[::ffff:7f00:1]` either by widening `RE_MAPPED_V4` or by classifying any `::ffff:` address as the embedded IPv4. 2. **Pass `redirect: "manual"` in `$fetch` defaults** and reject 3xx. (Compare `astro:assets`, which already does this — `await fetch(url, { redirect: "manual" })` and explicit 3xx-rejection.) 3. **Pin the validated IP to the connection.** Resolve the hostname once during validation, then pass a custom `undici.Agent` with `connect.lookup` returning the resolved IP only. This closes both the IPv6 bypass class (the resolved IP is checked again) and the redirect class (post-30x lookup is forced to the original IP). Reference: `request-filtering-agent` on npm. (2) alone closes Class 2. (1) alone closes Class 1. (3) closes both with one change.