/

CVE-2026-54745

Severity:

CRITICAL RISK

(10.0)

Unauthenticated SSRF (CWE-918) in Kubeflow Pipelines

CVE-2026-54745

Deferred

Vulnerability type

Category:

Server-Side Request Forgery

CWE:

CWE-284, CWE-918

Impact:

Unauthorized access to internal services and potential credential disclosure

Severity:

CRITICAL RISK

(10.0)

CVSS version:

3.1

Catch this CVE while code is being written

Detect vulnerable code directly in pull requests, so CVEs are fixed before merge.

Summary

Kubeflow Pipelines exposes an unauthenticated server-side request forgery (SSRF) condition through its frontend proxy routes, allowing attacker- controlled requests to reachable internal HTTP or HTTPS services. Depending on network access and service permissions, this may expose sensitive responses, cloud metadata credentials, or enable unauthorized changes to internal APIs.

Affected context

The issue occurs in Kubeflow Pipelines frontend deployments where the /_proxy/ route accepts attacker-controlled destinations without host allowlisting or filtering for loopback, link-local, RFC1918, or cluster-local addresses. The route may remain reachable without authentication or authorization enforcement, including through supported API prefixes and crafted Referer headers.

Exploit preconditions

  • Attack surface: Kubeflow Pipelines frontend proxy routes exposed over HTTP or HTTPS
  • Required access: Ability to send unauthenticated requests to a reachable frontend endpoint
  • Trigger point: Processing of a request through the /_proxy/ route or its API-prefixed equivalents
  • Why exploitation is possible: Attacker-controlled target origins are passed to the proxy without adequate destination validation, allowing requests to internal or otherwise reachable services

Why this happens

The underlying flaw is insufficient server-side validation of proxy destinations. The frontend accepts an arbitrary HTTP or HTTPS origin and forwards requests without enforcing a trusted host policy or excluding sensitive network ranges. This behavior maps primarily to CWE-918, Server-Side Request Forgery, and is compounded by CWE-284, Improper Access Control, because the proxy route can be reachable outside the intended authorization boundary.

Potential impact

  • Sensitive cloud instance metadata or workload credentials may be disclosed
  • Kubernetes, service, or cluster-internal API responses could be exposed
  • Internal services may receive unauthorized GET, POST, or other HTTP methods
  • Requests carrying forwarded headers or bodies could alter internal service behavior
  • An attacker could gain a foothold for credential misuse or lateral movement

General mitigation guidance

Mitigation generally centers on removing unauthenticated arbitrary proxying

and enforcing destination controls at the server boundary.

Common defensive patterns include:

  • Restricting proxy destinations to an explicit, minimal host and scheme allowlist
  • Rejecting loopback, link-local, RFC1918, RFC6598, IPv6 private, and cluster-local destinations
  • Resolving hostnames before connection and validating the resulting addresses
  • Ensuring proxy routes are covered by authentication and authorization middleware
  • Limiting forwarded headers, methods, request bodies, and response exposure
  • Applying network egress policies that prevent access to metadata and sensitive internal services
  • Running frontend components with least privilege and narrowly scoped service permissions
  • Avoiding proxy behavior that reflects upstream responses directly to untrusted clients

How teams detect issues like this

Detection approaches can include:

  • Static analysis for proxy middleware configured with user-controlled URLs
  • Code review checks for missing host allowlists, destination validation, and authorization coverage
  • Tests that attempt access to loopback, link-local, private, metadata, and cluster-local addresses
  • Policy checks confirming that proxy routes require authentication and enforce approved methods
  • Network monitoring for frontend-originated connections to metadata endpoints and unexpected internal services
  • Application logging and alerting on unusual Referer, target-origin, forwarding-header, and method combinations
  • Automated security review tools that identify SSRF sinks and missing access-control boundaries

Sources

SourceLink
github.comgithub.com/kubeflow/pipelines/commit/a35f97aa4c17b25572369e5022546ab4421bdbdd
github.comgithub.com/kubeflow/pipelines/pull/13511
github.comgithub.com/kubeflow/pipelines/releases/tag/2.17.0
github.comgithub.com/kubeflow/pipelines/security/advisories/GHSA-gqww-5pj5-8fq7

Frequently asked

questions

Yes. It can be remotely reachable when the Kubeflow Pipelines frontend is

exposed and the proxy route accepts unauthenticated requests. Exploitability

still depends on deployment routing, network reachability, and the services

accessible from the frontend.

The attacker generally needs only network access to a vulnerable frontend

endpoint. Authentication may not be required when the route is outside the

configured authorization boundary.

Remove arbitrary destination proxying or constrain it with strict host,

scheme, address-range, method, header, and authorization controls. Internal

network egress restrictions provide an additional defensive layer.

Treat all proxy destinations as untrusted input, enforce explicit allowlists,

test private and metadata address handling, require authorization on every

proxy route, and monitor frontend egress to sensitive internal services.

Find out what's already
exploitable in your codebase.