GPU VulnDB

Database/Container, Kubernetes & orchestration

Envoy: use-after-free in the HTTP ext_authz client crashes workers under production traffic

CVSS 5.9CVE-2026-50572Container, Kubernetes & orchestrationcurated

Impact

Envoy's HTTP external-authorization client can keep a stale request callback after a request has been rejected. When RawHttpClientImpl::onSuccess later processes the authorization response, it invokes callbacks_ on a destroyed owner - a use-after-free that crashes the process under ordinary production traffic. Where Envoy is the ingress or service-mesh sidecar in front of inference and control-plane services, a crash loop on the authz path means requests stop reaching GPU workloads even though the GPUs are healthy, and it hits exactly the deployments that enforce authorization centrally. The advisory establishes the trigger only for the HTTP ext_authz client, not for other filters. Memory corruption of this kind is not claimed to be exploitable beyond the crash in the record given.

Who can reach it

Reachable by unauthenticated network traffic through an Envoy configured with the HTTP ext_authz filter, where requests get rejected by the authorization service (CVSS AV:N/AC:H/PR:N).

What to do

Upgrade Envoy to 1.36.10, 1.37.6, 1.38.4 or 1.39.1 depending on your branch, and restart or roll the proxies - a control-plane-driven rolling restart of ingress gateways and sidecars, with no node drain. If you cannot patch now, the gRPC ext_authz client is not named as affected; switching the filter to it is a config change worth testing.

References

Related entries

All Container, Kubernetes & orchestration entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.