Astro: Unauthenticated path override in the @astrojs/vercel ISR function
When ISR is enabled, the serverless entrypoint lets an unauthenticated request
decide which route the origin renders. The internal _isr function reads the
x_astro_path query parameter and rewrites the request path to it without any
authentication. Edge level access controls only ever see the /_isr path, so
they do not apply to the route that actually gets rendered. This is the same
confused deputy problem as CVE-2026-33768, reachable again through the ISR path.
This affects apps that use @astrojs/vercel with isr: true and protect routes
at the edge. Two common setups are affected:
/admin).edgeMiddleware: true) where authorization lives in Astro
middleware, since that middleware runs at the edge and not in the origin.An attacker reads any GET rendered route by requesting
/_isr?x_astro_path=/the/protected/path. No credentials are required. The
protected content is produced by a fresh origin render, so the attack does not
depend on the response being cached first.
packages/integrations/vercel/src/serverless/entrypoint.ts picks the real path
like this:
if (hasValidMiddlewareSecret) {
realPath = request.headers.get(ASTRO_PATH_HEADER); // secret checked
} else if (request.headers.get('x-vercel-isr') === '1') {
realPath = url.searchParams.get(ASTRO_PATH_PARAM); // no secret checked
}
The header path is gated by the per build secret and is fine. The ISR branch is not. Two facts make it reachable by anyone:
_isr function is publicly addressable.x-vercel-isr: 1 on requests to it, including direct external
requests, so the attacker does not even need to send that header.So GET /_isr?x_astro_path=/admin sets the internal path to /admin and renders
it. The edge saw only /_isr, which is allowed, so any path based rule on
/admin never fires. In split deployments the edge middleware also runs against
/_isr, and the origin does not run middleware at all, so middleware based auth
is skipped as well.
CVE-2026-33768 was fixed in 10.0.2 by commit 335a204161 (PR #15959), which
required the secret for every path override and removed the query parameter
source. Commit aa266364fe (PR #16079, "Fix ISR path rewrite to prevent 404")
brought the query parameter back, guarded only by the x-vercel-isr header. That
header is not a security boundary, so the fix was effectively undone for ISR
routes starting in 10.0.3.
Worth noting the contrast: the original report treated Edge Middleware as the
mitigation and scoped the issue to deployments without it. Here, for split
deployments, Edge Middleware is bypassed too, since the attacker reaches /_isr
directly and the middleware only sees /_isr while the origin runs none.
output: 'server' and adapter
vercel({ isr: true })./admin that returns sensitive content./admin at the edge, for example a Vercel path rule that returns 403, or
a middleware auth check in a split (edgeMiddleware: true) build./admin. It is blocked (403)./_isr?x_astro_path=/admin. It returns 200 with the admin content.
The response header X-Vercel-Cache: MISS confirms it was rendered fresh, not
served from an existing cache entry./_isr.Unauthenticated read of any GET rendered route that is protected only at the edge. No integrity or availability impact because the vector is GET only.
One option would be to require the secret again for path overrides, the way
PR #15959 did, so the ISR branch stops trusting the client supplied x_astro_path.
The 404 that PR #16079 was fixing would then need another approach that does not
rely on client input.
왜 이 VPI인가 (설명가능 · 실험적)
VPI 산정 기준
| 영향도(기본값(정보 없음)) | 55.00 |
| 악용 신호(추가 악용신호 없음) | ×1.00 |
| VPI | 55.00 |
VPI 공식 vpi-v1 기준