hi guys,
commit: 40594bd98e6d6ed993b5c6021c93fdf96d2e5851 (as-of 2026-01-31) contact: GitHub Security Advisory (https://github.com/distribution/distribution/security/advisories/new)
in pull-through cache mode, distribution discovers token auth endpoints by parsing WWW-Authenticate challenges returned by the configured upstream registry. the realm URL from a bearer challenge is used without validating that it matches the upstream registry host. as a result, an attacker-controlled upstream (or an attacker with MitM position to the upstream) can cause distribution to send the configured upstream credentials via basic auth to an attacker-controlled realm URL.
this is the same vulnerability class as CVE-2020-15157 (containerd), but in distribution’s pull-through cache proxy auth flow.
HIGH
note: the baseline impact is credential disclosure of the configured upstream credentials. if a deployment uses broader credentials for upstream auth (for example cloud iam credentials), the downstream impact can be higher; i am not claiming this as default for all deployments.
credential exfiltration of the upstream authentication material configured for the pull-through cache.
attacker starting positions that make this realistic: - supply chain / configuration: an operator configures a proxy cache to use an upstream that becomes attacker-controlled (compromised registry, stale domain, or a malicious mirror) - network: MitM on the upstream connection in environments where the upstream is reachable over insecure transport or a compromised network path
registry/proxy/proxyauth.go:66-81 (getAuthURLs): extracts bearer realm from upstream WWW-Authenticate without validating destinationinternal/client/auth/session.go:485-510 (fetchToken): uses the realm URL directly for token fetchinternal/client/auth/session.go:429-434 (fetchTokenWithBasicAuth): sends credentials via basic auth to the realm URLattachment: poc.zip (local harness) with canonical and control runs.
the harness is local and does not contact a real registry: it uses two local HTTP servers (upstream + attacker token service) to demonstrate whether basic auth is sent to an attacker-chosen realm.
unzip -q -o poc.zip -d poc
cd poc
make canonical
make control
expected output (excerpt):
[CALLSITE_HIT]: getAuthURLs::configureAuth
[PROOF_MARKER]: basic_auth_sent=true realm_host=127.0.0.1 account_param=user authorization_prefix=Basic
control output (excerpt):
[CALLSITE_HIT]: getAuthURLs::configureAuth
[NC_MARKER]: realm_validation=PASS basic_auth_sent=false
validate that the token realm destination is within the intended trust boundary before associating credentials with it or sending any authentication to it. one conservative option is strict same-host binding: only accept a realm whose host matches the configured upstream host.
addendum.md poc.zip PR_DESCRIPTION.md RUNNABLE_POC.md
best, oleh
{
"cwe_ids": [
"CWE-918"
],
"severity": "HIGH",
"github_reviewed_at": "2026-04-06T17:52:52Z",
"nvd_published_at": "2026-04-06T15:17:10Z",
"github_reviewed": true
}