/

CVE-2026-92972

Severity:

HIGH RISK

(8.6)

Missing Authentication (CWE-306) in SGLang Routing

CVE-2026-92972

Deferred

Vulnerability type

Category:

Missing Authentication

CWE:

CWE-306

Impact:

KV Transfer Redirection, Metadata Disclosure, and Denial of Service

Severity:

HIGH RISK

(8.6)

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

SGLang's prefill bootstrap service exposes a routing endpoint without adequate authentication in prefill/decode disaggregation mode. An unauthenticated attacker may poison the KV transfer routing table, redirecting decode workers to attacker-controlled endpoints and potentially causing denial of service or exposing transfer metadata.

Affected context

The issue occurs in SGLang deployments using prefill/decode disaggregation mode when the prefill bootstrap service is network-reachable and its route management endpoint does not require effective authentication or equivalent access control. Risk depends on deployment exposure, network segmentation, and whether routing updates are trusted without validating their destination addresses.

Exploit preconditions

  • Attack surface: The prefill bootstrap service's HTTP route-management

endpoint

  • Required access: Network reachability to the endpoint; no valid credentials

may be required

  • Trigger point: Processing of a route update that supplies worker address

information

  • Why exploitation is possible: The service accepts attacker-controlled

routing values without first establishing the caller's identity and

authorization, allowing the KV transfer routing table to be altered

Why this happens

The underlying defect is missing authentication and authorization on a state-changing management endpoint. Under CWE-306, a security-sensitive function can be invoked without verifying that the requester is authorized. Because route updates influence where decode workers send KV transfer traffic, accepting unauthenticated destination data permits untrusted callers to influence internal service behavior and communication paths.

Potential impact

  • Attackers may redirect KV transfer traffic to unauthorized endpoints.
  • Decode workers could experience denial of service or failed inference requests.
  • KV transfer metadata, including session identifiers, may be disclosed.
  • Tensor-parallel topology parameters could be exposed.
  • Attackers could gain information useful for further targeting of the deployment.

General mitigation guidance

Mitigation generally focuses on protecting state-changing control-plane operations with strong authentication and authorization. Route updates can be restricted to trusted service identities, and destination addresses can be validated against an allowlist or service-discovery policy rather than accepting arbitrary values. Secure-by-default deployments should keep bootstrap and routing services on private network segments, reduce their privileges, and avoid exposing management endpoints beyond required peers. Additional integrity protections for routing updates and minimization of sensitive metadata can reduce residual impact.

How teams detect issues like this

Detection approaches include static and configuration analysis for HTTP handlers that mutate routing or service state without authentication middleware, as well as code-review checks for missing authorization around PUT, POST, or similar management operations. Runtime monitoring can alert on route changes from unexpected identities, source networks, or destination addresses. Policy enforcement may require authenticated service-to-service traffic, private management interfaces, and approved endpoint allowlists. Automated application-security review tools may help identify unauthenticated state-changing routes.

Sources

SourceLink
github.comgithub.com/sgl-project/sglang
github.comgithub.com/sgl-project/sglang/blob/v0.5.19/python/sglang/srt/disaggregation/common/conn.py#L1716-L1718
github.comgithub.com/sgl-project/sglang/blob/v0.5.19/python/sglang/srt/disaggregation/common/conn.py#L1736-L1810
github.comgithub.com/sgl-project/sglang/issues/39400
vulncheck.comvulncheck.com/advisories/sglang-through-0.5.19-unauthenticated-route-poisoning-via-put-endpoint

Frequently asked

questions

It may be remotely exploitable when the prefill bootstrap service is reachable

over a network and the route-management endpoint accepts unauthenticated

requests. Exposure is reduced when the service is isolated and reachable only

by trusted internal components.

The attacker generally needs network reachability to the affected endpoint.

Valid credentials may not be required, although network controls or service

mesh policies can still limit practical access.

Require authenticated and authorized service-to-service requests for routing

updates, validate destinations against trusted policies, and keep the

bootstrap service on a restricted network.

Treat every state-changing management endpoint as privileged, enforce

authentication and authorization by default, monitor routing-table changes,

and apply least-privilege and network-segmentation policies.

Find out what's already
exploitable in your codebase.