Advisory

GHSA-g6vg-wj8f-48cj

Details, part 2 of 2. The advisory

Details

Run the focused test with the project-supported Go toolchain:

go test ./internal/jwtverify -run TestJWKSCacheKeyIsNotScopedToTemplatedEndpointPoC -count=1 -v

Observed vulnerable output in my local test environment using Go 1.26.3:

=== RUN   TestJWKSCacheKeyIsNotScopedToTemplatedEndpointPoC
{"level":"info","endpoint":"http://127.0.0.1:32811/%7B%7Btenant%7D%7D/jwks.json","time":"2026-05-21T23:49:28+07:00","message":"JWKS manager created"}
{"level":"info","endpoint":"http://127.0.0.1:32811/%7B%7Btenant%7D%7D/jwks.json","time":"2026-05-21T23:49:28+07:00","message":"JWKS manager created"}
{"level":"info","endpoint":"http://127.0.0.1:32811/%7B%7Btenant%7D%7D/jwks.json","time":"2026-05-21T23:49:28+07:00","message":"JWKS manager created"}
--- PASS: TestJWKSCacheKeyIsNotScopedToTemplatedEndpointPoC (0.07s)
PASS
ok  	github.com/centrifugal/centrifugo/v6/internal/jwtverify	0.088s

The passing test demonstrates the vulnerable behavior because it asserts these controls:

  1. A legitimate tenant-B token signed by tenant B succeeds with a fresh verifier.
  2. A forged tenant-B token signed by tenant A fails with a fresh verifier.
  3. A legitimate tenant-A token succeeds and primes the JWKS cache with tenant A's shared-kid key.
  4. The forged tenant-B token signed by tenant A then succeeds with user ID victim.
  5. The tenant-B JWKS request counter does not increase during forged verification, proving the forged token was accepted from the cross-tenant cache hit rather than from tenant B's JWKS endpoint.

Expected behavior after a fix: the forged tenant-B token should remain rejected after tenant A primes the cache, or the verifier should fetch/consult tenant B's independent JWKS cache namespace before verification.

Impact

This is a cross-issuer / cross-tenant JWT authentication bypass in dynamic JWKS deployments.

Impacted deployments are those that use dynamic JWKS endpoint templates to select different JWKS URLs for different allowed issuers or audiences, for example multi-tenant deployments using {{tenant}} values extracted from iss or aud.

An attacker who can obtain or mint a valid token for one allowed issuer/tenant can authenticate as another allowed issuer/tenant if both JWKS documents use the same kid value and the attacker's issuer key is cached first. kid values are not globally unique by specification and are often operational labels such as current, default, or rotation identifiers, so the verifier should not rely on kid uniqueness across different JWKS trust domains.

Potential consequences include:

  • Authentication as a user in another issuer/tenant namespace.
  • Unauthorized connection-token acceptance.
  • Unauthorized subscription-token acceptance where separate subscription JWTs are configured.
  • Cross-tenant confidentiality and integrity impact when issuer-derived JWKS endpoints are used as separate trust domains.

Suggested remediation

Scope JWKS cache entries and singleflight keys to the resolved JWKS trust domain, not only to the JWT kid.

For dynamic endpoints, compute the endpoint namespace before cache lookup and use a composite cache key such as:

resolved_jwks_url + "\x00" + kid

or an equivalent canonical trust-domain identifier plus kid.

The same composite namespace should be used for:

  • TTL cache lookup.
  • TTL cache storage.
  • singleflight.Group.Do keys.

A regression test should prime tenant A's cache and then verify that a forged tenant-B token signed by tenant A remains rejected.

Sources & licences