You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
feat(kubernetes): establish outer-fence guarantees from a measured probe, not the NetworkPolicy object #3862
#3677 was the right call. networkPolicyEnforced was an operator claim with nothing behind it, and a flag that can be set on a cluster where nothing enforces reads like a check it is not. This asks for the option between that flag and a prerequisite: measure the fence.
The Kubernetes driver already fails closed on the policy object. validate_sandbox_runtime_fence compares the persisted spec, labels and annotations with the intended ones, and KubernetesOuterFenceEvidence::project then establishes DefaultDenyEgress, NoUnmanagedEgressPath, RevocationVerified and ControllerLossFailsClosed from the policy's UID and resourceVersion. All of that is control-plane state. The API server stores the same object when no CNI enforces it, so on that cluster the descriptor records all four guarantees and the sandbox starts.
The contract asks for more than that. from_enforcement_evidence "deliberately does not infer guarantees from the mere presence of evidence", and DefaultDenyEgress is defined as "No workload packet can leave without an explicit mediated decision". The Docker driver meets it by inspecting the running container, where the daemon that reports network_mode=none is the one enforcing it. Kubernetes has no equivalent, because the enforcer is the CNI and the API does not report on it.
Proposal, once per sandbox namespace before the boundary is confirmed:
After the workload fence is validated, have a pod under that policy attempt one connection to a destination the empty-egress policy must drop.
Project the four guarantees only if that attempt fails. Otherwise validate() rejects the boundary, as it already does when a guarantee is missing.
Cache the result per namespace, keyed to the policy UID and generation, so it is not a per-sandbox cost.
The open design question is the destination. It has to be reachable by the cluster network and dropped by the policy, which varies by cluster; a configurable target with a gateway-service default would work. docs/security/best-practices.mdx could carry the CNI caveat next to "enforce all network controls" in the meantime.
This issue has had no activity for 14 days and is now marked stale. It may be closed in 14 days if there is no further activity. Comment or remove the state:stale label to keep it open.
@drew re-checked at main c06ba109c. KubernetesOuterFenceEvidence::project (isolation.rs L42 to L59) still grants DefaultDenyEgress, RevocationVerified and ControllerLossFailsClosed from the stored policy's isolation flags and a zero egress rule count. Nothing measures the CNI yet, and #4067 binds the fence earlier without changing that.
The one call that blocks a probe PR is the drop target. Is a configurable destination that defaults to the gateway service acceptable? If so, the probe can follow against that default.
Activity
github-actions commentedon Oct 15, 2026
This issue has had no activity for 14 days and is now marked stale. It may be closed in 14 days if there is no further activity. Comment or remove the state:stale label to keep it open.
imran-siddique commentedon Oct 15, 2026
@drew re-checked at main
c06ba109c.KubernetesOuterFenceEvidence::project(isolation.rs L42 to L59) still grantsDefaultDenyEgress,RevocationVerifiedandControllerLossFailsClosedfrom the stored policy's isolation flags and a zero egress rule count. Nothing measures the CNI yet, and #4067 binds the fence earlier without changing that.The one call that blocks a probe PR is the drop target. Is a configurable destination that defaults to the gateway service acceptable? If so, the probe can follow against that default.