/

CVE-2026-35029

Severity:

HIGH RISK

(8.8)

Missing Authorization in LiteLLM AI Gateway

CVE-2026-35029

Modified

Vulnerability type

Category:

Missing Authorization

CWE:

CWE-863, CWE-425

Impact:

Remote Code Execution and Privilege Escalation

Severity:

HIGH RISK

(8.8)

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

LiteLLM's /config/update endpoint fails to enforce admin-level authorization, allowing any authenticated user to change critical server configuration and environment variables.

An attacker with basic platform access can leverage this to register custom handlers that execute attacker-controlled Python code, read arbitrary files, and take over other privileged accounts.

Affected context

This vulnerability affects LiteLLM deployments that expose the management or admin API to authenticated users who are not strictly isolated by role.

It is most relevant when:

  • The /config/update endpoint is reachable from user sessions intended only for normal proxy usage
  • Environment variables drive authentication, branding, or handler configuration
  • The same LiteLLM instance both serves end users and hosts administrative configuration endpoints

Exploit preconditions

  • Attack surface: HTTP(S) management endpoint /config/update on the LiteLLM proxy server.
  • Required access: Any authenticated LiteLLM user account; no admin or elevated role is required.
  • Trigger point: Sending crafted requests to /config/update to modify configuration and environment variables (e.g., handler paths, UI credentials, file paths).
  • Why exploitation is possible: The endpoint performs sensitive configuration updates but does not verify that the caller holds an administrative or privileged role, violating least-privilege and role-based access control expectations.

Why this happens

The underlying flaw is a missing authorization check on a highly sensitive configuration endpoint.

While authentication confirms the caller's identity, the system does not enforce that only administrators may invoke /config/update or restrict which fields non-admins can modify.

This aligns with CWE-863: Incorrect Authorization, where access control logic does not properly validate whether the authenticated principal is permitted to perform the requested operation.

In effect, configuration intended to be controlled only by trusted operators becomes writable by any logged-in user, enabling abuse of environment-driven behaviors such as dynamic handler registration and credential configuration.

Potential impact

  • Remote code execution by:
  • Registering custom pass-through endpoint handlers that load and execute attacker-controlled Python code
  • Arbitrary file read from the server by:
  • Setting UI_LOGO_PATH to sensitive file paths and retrieving the content via /get_image
  • Privilege escalation and account takeover by:
  • Overwriting UI_USERNAME and UI_PASSWORD to values controlled by the attacker
  • Persistent backdoor configuration that survives normal restarts if configuration is persisted
  • Lateral movement within the hosting environment if the compromised process has access to other internal services, credentials, or network segments

General mitigation guidance

Mitigations should focus on enforcing strict, role-aware access control around administrative functionality and hardening configuration flows:

  • Require explicit admin or operator roles to call configuration and management endpoints such as /config/update.
  • Centralize authorization logic so all privileged routes consistently enforce least-privilege checks.
  • Separate administration interfaces from user-facing proxy endpoints, ideally on distinct paths, ports, or network segments.
  • Limit which configuration parameters can be changed at runtime, and avoid allowing environment-variable overrides for security-sensitive values where possible.
  • Treat file path and handler registration settings as highly privileged operations and restrict them to trusted operators only.
  • Run the LiteLLM process with minimal OS-level privileges so that a compromise yields limited access to the filesystem and other resources.

How teams detect issues like this

Detection efforts can combine application-level monitoring with code-level review:

  • Static analysis and code review:
  • Search for routes that mutate configuration or environment-derived state and confirm they enforce role checks.
  • Flag any handler that updates sensitive fields (credentials, file paths, code handlers) without explicit authorization middleware.
  • Review how environment variables are read and updated to ensure they cannot be modified by unprivileged requests.
  • Runtime monitoring:
  • Log and alert on calls to /config/update, including caller identity, source IP, and changed fields.
  • Monitor changes to security-critical configuration values such as handler endpoints, UI credentials, and file path settings.
  • Track unexpected spikes in /get_image requests or anomalous file paths being requested via that endpoint.
  • Configuration governance:
  • Periodically audit effective configuration and environment settings to detect unauthorized changes.
  • Use automated compliance checks to ensure that management endpoints are restricted to administrative networks or roles.

Sources

SourceLink
github.comgithub.com/BerriAI/litellm/security/advisories/GHSA-53mr-6c8q-9789
seclists.orgseclists.org/fulldisclosure/2026/Apr/17
access.redhat.comaccess.redhat.com/errata/RHSA-2026:13545
access.redhat.comaccess.redhat.com/errata/RHSA-2026:28960
access.redhat.comaccess.redhat.com/errata/RHSA-2026:30056
access.redhat.comaccess.redhat.com/security/cve/CVE-2026-35029
bugzilla.redhat.combugzilla.redhat.com/show_bug.cgi?id=2455474
security.access.redhat.comsecurity.access.redhat.com/data/csaf/v2/vex/2026/cve-2026-35029.json

Frequently asked

questions

Yes, if the LiteLLM instance is exposed over the network and an attacker can obtain any valid user credentials, they can remotely call /config/update and potentially achieve code execution or privilege escalation.

The attacker must be an authenticated user of the LiteLLM platform, but does not need administrative privileges. Any account that can reach /config/update is sufficient under the vulnerable configuration.

Apply strict role-based access control to all configuration and management endpoints, ensuring only designated administrative roles can change sensitive settings. Additionally, restrict which configuration fields are mutable at runtime and isolate administrative APIs from general user traffic.

Design APIs with a clear separation between user and administrative operations, enforce centralized authorization checks, and treat configuration-changing endpoints as high-risk surfaces. Security reviews and automated tests should verify that every sensitive route requires the appropriate role or privilege before executing.

Find out what's already
exploitable in your codebase.