Patch
🔥 CyberForge CVE of the Day #099
🚨 CVE-2026-124440 — Kyverno namespace clamp bypass
Authenticated namespace tenant · classic Policy rights · controller authority
The namespace clamp can approve one target interpretation while apiCall reaches another. A valid controller identity does not establish that the tenant could direct it there.
Fix: Kyverno 1.19.2, through a complete supported upgrade. Verify serving controllers, supporting resources and intended policy enforcement.
🎯 The quick hit:
What happened? The namespace clamp added in 1.19.1 checks a decoded/cleaned copy of apiCall input. On a decoding error it can fall back to the raw representation, while the eventual request client interprets the input differently. The guard and the request can disagree about the protected target.
Who acts first? Owners of applicable Kyverno 1.19.1 installations where an authenticated namespace tenant can create classic namespaced Policies and cause their apiCall context to be evaluated. The admission-controller's actual roles determine the immediate reach.
Why this matters: WarlordSam07's maintainer-published advisory reports cross-namespace reads, controller-authorised admission-webhook creation and a full cluster-admin chain on the documented default lab. Secrets require controller read rights; the wider outcome needs functional admission and later useful workload/role authority.
First move: establish the serving engine, policy authoring and controller authority; preserve relevant history; arrange the supported correction. Follow author → policy/evaluation → controller request → target/result, then assess downstream effects separately.
Escalate: a corroborated unapproved cross-scope operation, webhook mutation or authority change. A controller username, a failed tenant request or a malformed string alone is not the finding.
🔑 Key details:
• Record / CNA: PUBLISHED / VulnCheck. This is a Kyverno policy-engine vulnerability, not a Kubernetes-core authentication bypass.
• CNA scope: >=1.19.1, <1.19.2; explicit 1.19.2 unaffected row. Classic namespaced http://kyverno.io/v1 Policy apiCall context.
• VulnCheck CNA CVSS v4.0 Base: 9.4 Critical — `CVSS:4.0/AV:N/AC:L/AT:N/PR:L/UI:N/VC:H/VI:H/VA:H/SC:H/SI:H/SA:H`.
• VulnCheck CNA CVSS v3.1 Base: 9.9 Critical — `CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H`.
• Maintainer: 9.9 v3.1, same vector; v4 vector/score null. Keep scoring versions and assessing authorities separate.
• NVD: Received; displays disclosure.com / Secondary v4 and Primary v3. “Primary” is a displayed provider role, not independent NIST authorship.
• Weakness: CNA CWE-441; maintainer also CWE-863 / CWE-918. Taxonomy is not proof of arbitrary external SSRF or unauthenticated access.
• Privileges / interaction: authenticated tenant Policy authoring and relevant evaluation; PR:L / UI:N.
• Fix: 1.19.2. Raw maintainer range >=1.19.1, patched 1.19.2; retain the CNA's bounded interval.
• CISA ADP/SSVC: absent in the captured CVE. No CISA “none” or active label is invented.
• KEV: no exact-ID match in the complete 1,769-entry catalog 2026.11.06, released 2026-11-06T19:44:45.1797Z. This is not proof of no exploitation.
• FIRST EPSS: successful exact-ID response, total: 0 / data: []. Probability, percentile and model date are unavailable, not zero.
• Credit: WarlordSam07, CNA finder / maintainer reporter. The exact CNA reference bridges its null-cve_id maintainer GHSA.
🧬 What actually went wrong?
A classic namespaced Policy can use apiCall context to ask the Kubernetes API for information or perform a supported operation. Kyverno runs the request with controller authority while its namespace clamp is meant to constrain where the policy author may direct it.
In 1.19.1 a decoding error can restore raw input for validation, while the original call proceeds to a client with different path/query handling. Guard and request can disagree about namespace. Blocking one example string cannot repair that representation class.
In the 1.19.2 tag, Fetch uses canonicalizeURLPath, validates the path-only result and assigns the canonical path back to the call before Execute. The helper normalises decoded Path, clears RawPath and reconstructs the request URI; query content does not satisfy namespace validation. Absolute URL and fragment forms are rejected, with separate target guards in the production call path.
Selected tag files, PR metadata and regression source were read passively. The final tag uses namespaceFromPath; an intermediate diff is not identical final code. No build, regression execution or binary attestation was performed.
📦 Which deployments belong in scope?
Establish these conditions together:
1⃣. An applicable 1.19.1 Kyverno engine actually evaluates the classic namespaced Policy apiCall path.
2⃣. A tenant has the relevant authenticated API access and create on http://policies.kyverno.io in its namespace; a namespace account alone is insufficient.
3⃣. A policy/context evaluation can occur. An eventual denial or Audit-mode policy is not proof that an apiCall side effect never ran.
4⃣. The admission-controller's effective RBAC can reach the target operation. An attempted read is not universal Secret access.
5⃣. For the wider lab chain, a functional webhook and suitable later admission/workload/role authority must exist and be exercised.
Reported lab: Kubernetes 1.31.0, Kyverno 1.19.1, stock Helm chart 3.9.1, default values; not all deployments tested.
Minimum correction: 1.19.2. Confirm the product/distributor build, all active admission instances and supporting CRDs/RBAC. Earlier releases are outside this CNA interval, not globally certified safe; custom backports and prereleases need their own evidence.
Migration caveat: today's upgrade guide also covers 1.20, where legacy Policy APIs remain readable but the policies no longer execute. A retained object or migration acknowledgement is not proof of replacement enforcement. Plan the version-specific supported path and test intended protection; “upgrade to latest” is incomplete guidance.
🕰️ Exposure window and timeline:
• 20th October, 23:50:19 UTC: stable v1.19.2 published; not draft or prerelease.
• 23rd October, 07:37:37 UTC: maintainer advisory published/updated; CNA datePublic is also 23rd October.
• 8th November, 13:15:33.723 UTC: CVE published/updated. NVD published/modified 14:18:10.633 UTC, status Received.
• 9th November: CyberForge checks to the stated cutoff. The newer identifier does not make the fix a new 9th November release.
Reconstruct the local window from serving build, delegated rights, evaluation and correction history. Document earliest reliable retention, clock offsets and collection gaps. Disclosure is not first exploitation. Current rights or an empty object list cannot clear a deleted policy or older state.
🗺️ MITRE ATT&CK:
Conditional capability mapping: T1068 — Exploitation for Privilege Escalation, Privilege Escalation / TA0004, Enterprise / Containers context. The official page lists Containers; checked v1.6, modified 24th October 2025, on 9th November.
Behaviour and basis: the maintainer reports an authenticated policy author reaching more powerful controller/cluster authority through the namespace-boundary flaw. The local evidence question is whether an applicable policy/evaluation induced a protected operation and later actual authority use.
Defender pivot: policy/permission history, audit/controller records and actual target/authority effects. MITRE defines T1068; the advisory supplies its case basis. A CWE or score does not establish an incident.
Limits: no field actor/campaign, node escape, persistence or exfiltration established. This authenticated-tenant route does not automatically map to public-facing initial access. No detector or coverage test.
🧭 SOC investigation trail:
Objective: establish whether tenant policy evaluation induced a controller operation beyond its intended scope, what resulted and whether later admission/authority effects occurred.
Collection: define clusters, UTC interval, audit policy/levels and collection health. Audit Event can supply user, verb, objectRef, requestURI, stage, responseStatus and timestamps. Metadata-only logs may omit policy bodies. Keep sensitive records in the authorised evidence store.
C01 — Establish applicability and reachability
Action / location: check serving admission Pods, images/digests, install history, active classic Policy APIs, tenant authoring and actual controller roles. Include every serving instance.
Signal / correlate: cluster, namespace, workload/Pod UID, image, policy API, rights and UTC; distinguish author and controller.
Benign causes / limits: stale/inactive images, admin-only authoring and evidenced backports alter applicability. A classic object listable on 1.20 can be unenforced.
Escalate: an applicable uncorrected delegated-policy deployment to its platform owner; local exposure is unassessed here.
C02 — Review the policy and evaluation
Action / location: preserve classic Policy specification/version/deletion history, historical authorisation and the admission activity that would evaluate its context. Do not replay the policy.
Signal / correlate: audit author, operation, target, response/stage and UTC; policy UID/resourceVersion and a plausible trigger. Different requests can have different auditIDs.
Benign causes / limits: GitOps and policy development look similar; missing bodies hide intent. Audit mode or an eventual denial does not prove that an apiCall had no effect.
Escalate: an unexplained policy/evaluation whose protected target conflicts with authorised tenant scope or intent.
C03 — Establish the controller operation and result
Action / location: review controller logs and API audit for cross-namespace reads or cluster-scoped webhook actions under the actual admission-controller account and roles.
Signal / correlate: policy/evaluation → controller user/operation → target namespace/resource/name → response and independent result. Use real joins; no universal auditID is assumed.
Benign causes / limits: controllers normally read across namespaces and manage webhooks. Username or managedFields.manager alone is not causal attribution; success alone does not describe disclosed data.
Escalate: a corroborated unapproved cross-scope operation with its exact result and data scope established.
C04 — Review admission configuration and effects
Action / location: preserve webhook before/after state, author/time, rules, selectors, clientConfig and failure handling against approved change records. Do not probe its endpoint.
Signal / correlate: controller write → actual match scope/backend → eligible admission → stored object mutation and version.
Benign causes / limits: chart/certificate/GitOps changes are alternatives. Registration alone proves neither a functional backend nor cluster-admin access.
Escalate: corroborated unapproved registration or mutation, balancing evidence with active harm and admission availability.
C05 — Review supported follow-on authority changes
Action / location: where admission abuse is supported, review workload/container/service-account changes, resulting identity use and Role/ClusterRole or binding edits. Include widened existing roles.
Signal / correlate: mutated workload UID/specification → running identity → later API operation/role result; establish the extra authority conditions of any wider chain.
Benign causes / limits: role aggregation and approved rollouts are alternatives. Token exposure, node-root or persistence is not supplied by a webhook alert alone.
Escalate: corroborated unauthorised identity/role effects linked to the earlier operation; scope response to supported evidence.
C06 — Verify correction and remaining questions
Action / location: owner verifies complete 1.19.2 correction, all active admission images/Pods, CRDs/RBAC, health and intended benign policy enforcement. A later 1.20 move needs migration acceptance.
Signal / correlate: release/change record, rollout UTC, active instances, API support and protected behaviour; keep historical incident disposition separate.
Benign causes / limits: changed tags, terminated Pods, retained objects or failed probes do not certify acceptance. Rollback can restore the vulnerable state.
Escalate: residual affected instances, failed enforcement/health or continuing unexplained operations. All six checks are proposed; not run.
The analyst handover:
Pass cluster/time scope, author/policy/evaluation, controller, target/result, corroboration, alternatives, actions, gaps and next owner. A no-match needs assets, interval, filters, retention and collection health. No findings ≠ proof of a clean system.
🔎 Indicators and observable behaviour:
No validated field IOC set established. Synthetic advisory names, markers and endpoints are demonstration material, not actor infrastructure.
Leads: unexplained policy context, cross-scope controller calls, unapproved webhook mutations and related workload/role edits. Encoding errors can be benign; presence is not proof and absence is not clearance.
The source reports cross-namespace ConfigMap reading and a functional/full lab chain. It makes Secret reads conditional on controller privileges and expressly leaves the PolicyException write not separately exercised. Preserve those evidence levels. Public examples were read passively; a fuller helper bundle is offered to maintainers, with its complete public availability unverified.
🧰 Optional CVE checks:
Documentation checked / POSIX syntax parsed only. No kubectl, API request or product test was run. Use approved contexts and scope; the placeholders deliberately stop accidental use.
Check A — What controller images and Pods are serving?
Platform owner: approved get/list rights, verified namespace and compatible kubectl/server versions:
```sh
kube_context='REPLACE_WITH_APPROVED_CONTEXT'
kyverno_namespace='REPLACE_WITH_KYVERNO_NAMESPACE'
case "$kube_context:$kyverno_namespace" in
*REPLACE_WITH*) printf '%s\n' 'Set approved context and namespace first.' >&2; exit 64 ;;
esac
kubectl --context "$kube_context" --namespace "$kyverno_namespace" \
get deployments,pods -o json
```
Inspect image/imageID, readiness, service-account and active instances. Correlate install/provenance and CRDs/RBAC; a tag is insufficient. Access/API errors are failed collection. Keep output in the approved store.
Check B — Can the intended tenant create classic Policies?
Use an approved real tenant context; owner confirms identity/groups/bindings. An administrator context answers a different subject's question.
```sh
tenant_context='REPLACE_WITH_APPROVED_TENANT_CONTEXT'
tenant_namespace='REPLACE_WITH_TENANT_NAMESPACE'
case "$tenant_context:$tenant_namespace" in
*REPLACE_WITH*) printf '%s\n' 'Set approved tenant context and namespace first.' >&2; exit 64 ;;
esac
kubectl --context "$tenant_context" auth can-i create http://policies.kyverno.io \
--namespace "$tenant_namespace"
```
Yes/no concerns this subject/context/time. Errors are not no. The check executes no Policy and clears no history. Controller target rights need separate RBAC evidence; do not add privileges to run this.
🩹 Remediation and verification:
Permanent correction: complete supported Kyverno 1.19.2, or obtain independently evidenced distributor/backport guidance. The official guide warns that an image-tag bump alone is insufficient: supporting resources, including CRDs, change. Plan controllers, RBAC, installation method and availability together.
Follow version-specific guidance. YAML uninstall/reinstall and CRD deletion can affect stored policies. A move to 1.20 needs legacy-policy migration and proof of replacement enforcement; a retained object or chart record is insufficient.
If delayed: platform owners can assess limiting untrusted authoring or excess controller authority while preserving required functions. These are analyst risk-reduction proposals, not certified workaround or patch equivalence. A literal input filter does not close the representation class.
Acceptance: all active admission images/digests, CRDs/RBAC, intended benign enforcement, service health, rollout UTC and owner. Investigate prior reads, webhooks and authority changes separately; updating does not remove already-created objects or settle past access. Recovery/credential work needs its own evidence. No change was performed here.
🧾 Evidence separation:
Verified source facts: published CNA identity/range/scores; maintainer patched version; dated catalog/forecast responses and official documentation.
Attributed research: WarlordSam07's documented lab, controls and full-chain observations, published in the maintainer advisory. Not independently reproduced; no field activity verified in this scoped review.
CyberForge interpretation: conditional priority, T1068 capability lens, SOC questions, read-only examples, acceptance criteria and handover design.
Local outcomes: exposure, compromise, remediation, detector performance and recovery are unassessed. Source reading, shell parsing and artifact review are the work actually performed.
🔥 CyberForge verdict:
Applicable 1.19.1 deployments that delegate classic Policy creation deserve prompt platform action and focused historical review. The important decision is which tenant can direct which controller authority to which protected target—not whether an object happens to carry a namespace label.
Establish the prerequisites, preserve useful history, complete the supported correction and prove intended enforcement. Then resolve suspicious prior operations with their own evidence and owners. An unavailable EPSS score, a KEV non-match or a quiet present-day log cannot provide that closure.
Validate the target you will use. Trace the authority you actually used.
🔗 Sources and further reading:
Maintainer advisory / researcher credit: https://github.com/kyverno/kyverno/security/advisories/GHSA-39w8-hv5g-f73w
CVE Program: https://cveawg.mitre.org/api/cve/CVE-2026-124440
Kyverno 1.19.2: https://github.com/kyverno/kyverno/releases/tag/v1.19.2
CISA KEV feed: https://www.cisa.gov/sites/default/files/feeds/known_exploited_vulnerabilities.json
FIRST exact-ID query: https://api.first.org/data/v1/epss?cve=CVE-2026-124440
Kubernetes auditing: https://kubernetes.io/docs/tasks/debug/debug-cluster/audit