GHSA-qf5v-m7p4-95rp

Fission v1.24.0 added PodSpec safety validation for tenant-facing Environment and Function CRDs (ValidatePodSpecSafety / ValidateContainerSafety admission webhook + sanitizeContainerSecurityContext executor merge layer), but the capability check was implemented as a fixed denylist of six Linux capabilities (SYS_ADMIN, NET_ADMIN, SYS_PTRACE, SYS_MODULE, DAC_READ_SEARCH, DAC_OVERRIDE). The denylist omitted CAP_SYS_TIME, among others. As a result, a tenant who could create a Function or Environment CRD could request securityContext.capabilities.add: ["SYS_TIME"], pass Fission's admission validation and merge-layer sanitization, and run attacker-controlled code with CAP_SYS_TIME in the resulting function or runtime container.

Demonstrated consequence: cross-tenant node integrity damage via CAP_SYS_TIME. The Linux real-time clock is not namespaced — time namespaces virtualize only MONOTONIC and BOOTTIME, never REALTIME — so a tenant container holding CAP_SYS_TIME could call clock_settime(CLOCK_REALTIME) and rewrite the shared node wall clock. That corrupts TLS / certificate validity windows, Kubernetes lease renewal, token expiry, scheduling, and time-series for every workload on the node.

The denylist also omitted SYS_RAWIO, BPF, SYS_RESOURCE, and MAC_ADMIN. Those are documented as evidence that the denylist is structurally incomplete (their practical impact is kernel-, LSM-, or device-cgroup-dependent and is not exercised in this report).

The deeper structural problem: a denylist on capabilities.add cannot constrain capabilities the OCI runtime grants by default — DAC_OVERRIDE is in the OCI default cap set and reaches the container regardless of any add check, partially mooting the denylist for its own entries. A capability allowlist (with drop:["ALL"] to remove the default set) is the only model that addresses both problems.

Affected

• Project: github.com/fission/fission
• Versions: <= 1.24.0
• Audited commits: v1.24.0 tag (ce617120) and current HEAD at audit time
• Component: pkg/apis/core/v1/podspec_safety.go (dangerousCapabilities) and pkg/executor/util/merge.go (dangerousMergeContainerCapabilities)
• Pre-condition: cluster does not enforce a restrictive Pod Security Admission (PSA) profile on the function-pod namespace. With PSA restricted in force the API server's PodSecurity admission rejects the pod at creation and this finding does not apply. Fission added its own PodSpec validation precisely because it cannot assume PSA enforcement.

Fix section (paste into the Fix / Patches field)

Fixed in v1.25.0 by:

• PR #3465 (commit 2569b42b) — replace the denylist with a PSA-restricted allowlist (NET_BIND_SERVICE only) at both enforcement layers:
  • pkg/apis/core/v1/podspec_safety.go — ValidateContainerSafety now rejects any capabilities.add entry not in allowedCapabilities. The container check is invoked from ValidatePodSpecSafety for every PodSpec container / init-container and from Environment.validateForAdmission for the bare Runtime.Container / Builder.Container.
  • pkg/executor/util/merge.go — sanitizeContainerSecurityContext filters capabilities.add through the same allowlist at every merge site (poolmgr / newdeploy / container executor / builder).
• The same PR adds 21 bounded CEL XValidation rules covering the cheap pod-level invariants (hostNetwork / hostPID / hostIPC / serviceAccountName / serviceAccount override) on FunctionSpec, Runtime, and Builder, plus the bare-Container SecurityContext checks (privileged != true, allowPrivilegeEscalation != true, capabilities.add ⊆ {NET_BIND_SERVICE}) on Runtime and Builder. The API server now short-circuits those attack vectors per CRD apply before the webhook is invoked. Per-container PodSpec iteration stays in the webhook because it exceeds the API server's CEL cost budget.
• The same PR also adds two forward-compat regression guards in pkg/apis/core/v1/podspec_safety_test.go:
  • TestAllTenantContainerSurfacesAreValidated walks every CRD root type via reflection and fails if any reachable *apiv1.PodSpec or *apiv1.Container field is missing from the hand-maintained known-covered set.
  • TestTenantContainerSurfaces_RejectSysAdmin end-to-end exercises each covered surface and asserts ValidateForAdmission rejects a SYS_ADMIN injection at that exact path.

Not addressed in v1.25.0

The advisory's structural recommendation to force capabilities.drop: ["ALL"] at the merge layer is not part of this fix. Fission's own sidecar containers (fission-fetcher, builder) were authored against the OCI default capability set and need a per-container cap audit before drop:["ALL"] can be applied uniformly. The allowlist on capabilities.add closes the demonstrated CAP_SYS_TIME impact; the OCI-default-cap concern (which the advisory itself marks as conditional / not demonstrated) is tracked separately.