[email protected]

LiteLLM Vulnerability GHSA-7hp6-4w63-5g45: Check Versions and Harden the Gateway

LiteLLM Vulnerability GHSA-7hp6-4w63-5g45: Check Versions and Harden the Gateway

Editorial note: The information in this article was compiled to the best of our knowledge at the time of publication. Technical details, prices, versions, licensing terms, and external content may change. Please verify the information provided independently, particularly before making business-critical or security-related decisions. This article does not replace individual professional, legal, or tax advice.

Running LiteLLM in your organisation? WZ-IT checks, updates and hardens LiteLLM gateways and takes over operation with CVE monitoring, see Managed LiteLLM and CVE monitoring. Book a meeting

On 30 September 2026, BerriAI published a security advisory for the LiteLLM proxy: GHSA-7hp6-4w63-5g45. An authenticated user with the internal_user role can escalate to proxy_admin and then execute arbitrary commands on the host through the MCP stdio endpoint. GitHub rates the flaw CVSS 9.9.

In many organisations, LiteLLM is the central access layer for language models: it bundles local models and provider APIs behind an OpenAI-compatible endpoint, manages keys and budgets, and holds the credentials of every connected provider for that purpose. Whoever takes over the gateway gains access to exactly those credentials. What LiteLLM does and where it sits in the stack is explained in What is LiteLLM?.

This article summarises how the flaw works, which versions are affected and patched, what the workaround costs in functionality, and which checklist secures a LiteLLM gateway beyond this case. All version information reflects 1 October 2026.

Table of Contents

  1. The vulnerability at a glance
  2. How the privilege escalation works
  3. Affected and patched versions
  4. Workaround if the update has to wait
  5. What to check after possible misuse
  6. LiteLLM vulnerabilities in 2026
  7. Checklist: hardening a LiteLLM gateway
  8. Our approach at WZ-IT
  9. Further guides

The vulnerability at a glance

Attribute Value
Identifier GHSA-7hp6-4w63-5g45
CVE none (as of 1 October 2026)
Title Privilege Escalation to Proxy Admin via Cross-Domain Reuse of the Salt Key
Published 30 September 2026
Severity critical, CVSS 3.1 9.9
CVSS vector AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H
CWE CWE-269, CWE-345, CWE-441
Prerequisite authenticated account with the internal_user role
Impact proxy_admin privileges, command execution on the host via MCP stdio
Package litellm (PyPI), proxy deployments
Reported by Hoa X. Nguyen (OPSWAT Unit 515)

Source: GitHub Security Advisory GHSA-7hp6-4w63-5g45.

The vector describes a flaw that is exploitable over the network without user interaction and requires only low privileges. S:C (scope changed) means the impact extends beyond the attacked component: a user account in the gateway turns into code execution on the server.

How the privilege escalation works

The core of the problem is one key with two jobs. According to the advisory, LiteLLM uses the same key to encrypt stored secrets and to mint session tokens. The documentation calls this key LITELLM_SALT_KEY; among other things it encrypts the model provider credentials stored in the database (LiteLLM, Production Best Practices).

The sequence according to the advisory:

  1. The attacker signs in with an account that has the internal_user role.
  2. They request a new API key and place a forged admin credential in a metadata field as the supposed secret.
  3. The proxy encrypts this value with the shared key and returns the result.
  4. The attacker presents the encrypted value as a bearer token.
  5. The proxy decrypts the token, trusts the forged admin identity and grants full administrative access.
  6. As proxy_admin, the attacker can execute commands on the host through the MCP stdio endpoint.

This is a classic confused deputy pattern (CWE-441): the proxy becomes the tool that produces a valid token for the attacker. Nothing in the encryption itself is broken. The weakness is that data from two trust domains is protected with the same key and the proxy does not distinguish what a decrypted value was meant for.

proxy_admin is the highest role in LiteLLM, with full control over organisations, teams and users. internal_user can sign in, view their own keys and, depending on team permissions, create their own keys (LiteLLM, Access Control). The gap between these two roles is what the flaw bridges.

The MCP stdio transport starts a local process using command and args from the configuration (LiteLLM, MCP). That is intended for MCP servers, but it turns any ability to add or test stdio servers into command execution on the gateway host. Why MCP servers should generally be treated as code with their own privileges is covered in the article on the Model Context Protocol in organisations.

Affected and patched versions

The advisory distinguishes two cases: from version 1.91.0, the vulnerable path is active in the default configuration. In versions 1.87.0 to 1.90.x it is only active if EXPERIMENTAL_UI_LOGIN=true was explicitly set.

Affected range (PyPI litellm) Patched version Release
1.87.0 to 1.90.x only affected with EXPERIMENTAL_UI_LOGIN=true; no separate fix in these lines -
1.91.0 to below 1.100.4 1.100.4 30 Sep 2026
1.101.0 to below 1.101.3 1.101.3 30 Sep 2026
1.102.0 to below 1.102.2 1.102.2 30 Sep 2026
1.103.0 to below 1.103.1 1.103.1 30 Sep 2026
1.104.0rc1 1.104.0rc2 (pre-release) 30 Sep 2026

Sources: version ranges from the advisory, release dates from the LiteLLM releases. On 1 October 2026, 1.101.4 and 1.103.2 followed as further releases in these lines.

Three points that are easy to miss in practice:

  • No backports for 1.91 to 1.99. Anyone running a version from these lines will not find a fix there. The affected range extends to below 1.100.4. This also applies to lines that received security updates for other flaws in August 2026.
  • Docker images follow the same numbering. The official images are published at ghcr.io/berriai/litellm with tags such as v1.100.4. A tag such as main-latest says little about which version is actually running.
  • Checking the version. The running version is shown by pip show litellm in the Python environment or by the litellm_version field in the response of /health/readiness. For containers, the image tag actually started is what counts, not the one intended in a compose file.

Versions before 1.87.0 are not affected by this flaw. They are, however, affected by other, partly actively exploited flaws (see LiteLLM vulnerabilities in 2026) and are not a safe state.

A second advisory appeared on the same day: GHSA-hhww-mrg2-969h, a reflected cross-site scripting flaw in /sso/debug/callback (CVSS 6.8) in versions 1.65.5 to below 1.85.0. Updating to one of the patched versions above also resolves it.

Update of 12 October 2026: On 1 October 2026 another advisory with a similar impact followed: GHSA-g5ff-637f-6q2m (CVSS 8.1, no CVE number). Even the internal_user_viewer role can read arbitrary files of the proxy process by sending a crafted vertex_ai_credentials value to /utils/transform_request, for example /proc/self/environ with the master key and provider keys. All versions below 1.95.0 are affected; anyone who updated to one of the patched versions above is not affected. GHSA-7hp6-4w63-5g45 still has no CVE number and is not listed in the CISA KEV catalog. Current stable releases are 1.105.0 and 1.104.3 from 11 October 2026, plus 1.103.4 (7 October), 1.102.4 and 1.101.6 (8 October).

Workaround if the update has to wait

The advisory names one workaround: set the environment variable EXPERIMENTAL_UI_LOGIN=false. This disables the vulnerable authentication path. In LiteLLM's token check, this path is only skipped when the variable is explicitly set to false (litellm/proxy/auth/user_api_key_auth.py). A missing entry is not enough.

Aspect Effect of EXPERIMENTAL_UI_LOGIN=false
Vulnerable authentication path disabled
CLI SSO login stops working
Claude Code gateway login stops working
API access with virtual keys (sk-...) remains possible, separate authentication path
Replacement for the update no

Teams that connect developer tools through the gateway should therefore plan the workaround as a short bridge: set it, restart the proxy, inform users, test and roll out the update, then reassess the variable deliberately.

What to check after possible misuse

The update closes the flaw but does not undo exploitation that has already happened. Whether it happened cannot be answered across the board. The review depends on who had access to accounts with the internal_user role and how long an affected version was running.

Check What to look for
Users and roles new or promoted proxy_admin accounts, unknown teams or organisations
API keys keys with unusual metadata, new keys with high budgets or without model restrictions
MCP servers newly added or modified stdio servers, unknown command entries
Host unknown processes, files and outbound connections of the proxy container or host
Provider accounts unusual consumption or access in the model providers' consoles
Logs access to admin endpoints by accounts that are not admins

If misuse cannot be ruled out, clean-up includes:

  • Rotate provider keys. A proxy_admin can use the stored model credentials, and command execution on the host also gives access to environment variables.
  • Rotate the master key. LITELLM_MASTER_KEY is the admin credential for the API and the UI; LiteLLM documents a dedicated rotation flow for it (Production Best Practices).
  • Change the salt key only with a plan. The documentation warns that changing LITELLM_SALT_KEY makes already encrypted credentials unreadable. A change therefore means re-entering all model credentials.
  • Rebuild the host. If there are signs of command execution, a cleanly rebuilt container or host is more reliable than hunting for every single change.

Whether an incident has to be reported under GDPR or NIS2 depends on which data passed through the gateway and which systems were reachable from the host.

LiteLLM vulnerabilities in 2026

GHSA-7hp6-4w63-5g45 is not the first critical LiteLLM flaw this year. Three earlier flaws are listed in the Known Exploited Vulnerabilities catalog of the US agency CISA, meaning exploitation has been confirmed.

Identifier Type Patched from In CISA KEV since
CVE-2026-42208 (GHSA-r75f-5x8p-qvmc) SQL injection in API key verification 1.83.7 8 May 2026
CVE-2026-42271 (GHSA-v4p8-mg3p-g94g) command execution via MCP stdio test endpoints, including with an internal_user key 1.83.7 8 Jun 2026
CVE-2026-59822 (GHSA-7488-6r32-c95q) MCP authentication bypass via OAuth2 passthrough 1.84.0 2 Sep 2026

As of 1 October 2026; BerriAI maintains the full list under Security Advisories. In addition, there was the supply chain incident of 24 March 2026: the PyPI packages litellm 1.82.7 and 1.82.8 were compromised for around 40 minutes, triggered by a compromised Trivy dependency in the CI pipeline. According to BerriAI, the official Docker image was not affected (LiteLLM, Security Update March 2026).

Two conclusions for operations follow from this:

  • MCP is the recurring attack surface. Two of the three exploited flaws and the current flaw involve MCP. Disabling MCP in the gateway when it is not needed reduces the attack surface considerably.
  • The release cadence is high. LiteLLM publishes several releases per week and maintains several stable lines in parallel. Without a fixed update process, a gateway drops out of the patched range within a few weeks. How structured vulnerability monitoring for self-hosted software is set up is described in CVE monitoring for self-hosted software.

The Ollama flaw Bleeding Llama showed a similar pattern in spring: self-hosted AI components are only as secure as their operation. For n8n, which is often combined with LiteLLM, the article on the n8n security update September 2026 summarises the current advisories.

Checklist: hardening a LiteLLM gateway

The following checklist goes beyond the current flaw. It is aimed at operators who use LiteLLM as the central gateway for several teams or applications.

Area Measure Relation to the current flaw
Version pin the version (image tag or package version), apply updates deliberately after testing direct: update to a patched version
Version verify the Docker image with cosign against the published key before starting (release notes) supply chain
Monitoring track advisories by GHSA identifier, not only by CVE current flaw has no CVE
Network do not expose the proxy directly to the internet; allow access only from defined network segments or via VPN narrows the circle of possible attackers
Network make the admin UI and management endpoints reachable only from the admin network; DISABLE_ADMIN_UI=true if the UI is not needed admin paths
Network restrict the gateway's outbound connections to the model providers in use makes data exfiltration after takeover harder
Roles keep proxy_admin accounts to a minimum; do not create internal_user accounts for people without a need prerequisite of the attack
Roles give applications virtual keys with model and budget limits instead of user accounts limits the damage of individual keys
Keys master key and salt key separate, randomly generated and stored in a secret manager; do not change the salt key after the first model is added protection of stored credentials
MCP disable MCP features if unused; add stdio servers only via configuration file and with approval command execution via MCP stdio
Host run the proxy as an unprivileged user in a container, without the Docker socket, with minimal environment variables limits the consequences of command execution
Logging log usage, admin actions and key changes centrally, for example with Langfuse for model calls detection
Experimental set experimental switches such as EXPERIMENTAL_UI_LOGIN only deliberately and document them root cause of the flaw

On the network segment: the common assumption that a gateway on the internal network is protected does not hold for this flaw. The attacker needs a valid user account, so they usually come from the internal network or through a compromised account. Network segmentation reduces the attack surface but replaces neither the update nor tight role assignment.

On MCP: an MCP server using the stdio transport is a process on the gateway host. Teams that provide MCP tools for agents should rather run them as separate services with HTTP transport in a separate segment and only allow them in the gateway, instead of starting arbitrary commands on the gateway itself. How agents work with limited permissions and approval steps is described in AI agents: permissions and approvals.

On provider choice: the gateway is also where you define which models are reachable at all. How European model APIs and your own inference are combined through LiteLLM is shown in the comparison of European LLM APIs.

Our approach at WZ-IT

WZ-IT operates LiteLLM as Managed LiteLLM with TLS, network segmentation, monitoring and fixed operating processes. For an advisory such as GHSA-7hp6-4w63-5g45, the work runs in five steps:

  1. Determine exposure. Record the running version, image tag, environment variables and active login paths and compare them against the version ranges in the advisory.
  2. Mitigate until the update. If needed, set EXPERIMENTAL_UI_LOGIN=false and coordinate the effect on CLI logins with users.
  3. Apply the update. Check the patched version in a test environment, test routing, fallbacks, budgets and integrations, then roll it out to production.
  4. Look for traces. Review roles, keys, MCP entries and logs for anomalies; rotate keys in a controlled way if misuse is suspected.
  5. Harden and monitor continuously. Implement the checklist above and add the gateway to CVE monitoring, which evaluates vendor security advisories in addition to the NVD.

Alongside your own inference on the AI Cube or a managed GPU server from WZ-IT, the gateway can also connect European model APIs. Support, consulting and implementation by WZ-IT.

Further guides

Have your LiteLLM version checked? We compare your installation against the advisory, apply the update and harden the gateway according to the checklist. Book a meeting

Sources

Timo Wevelsiep

Written by

Timo Wevelsiep

Co-Founder & CEO

Co-Founder of WZ-IT. Specialized in cloud infrastructure, open-source platforms and managed services for SMEs and enterprise clients worldwide.

Frequently Asked Questions

Answers to important questions about this topic

Customers from mid-sized business, industry and the public sector, in Germany and internationally

  • Stadtwerke Brühl
  • DGHO e.V.
  • solarwerkstatt berlin
  • Krüger Dirndl
  • ABCO Water Systems
  • Golem.de
  • EVADXB
  • nextGYM
  • AInergy
  • ml&s
  • Odiseo Solutions
  • Annota
  • ARGE
  • SweetConnect GmbH
  • Aphy AG
  • CORGOS
  • Rekorder
  • SolidProof
  • Yonju
  • Keymate
  • Paritel
  • Mr. Clipart
  • Millenium
  • Negosh
  • Führerscheinmacher
  • BTI
  • Deutsche Gesellschaft für Datenschutz
Read client reviews

Enquiry

Check, update and harden your LiteLLM gateway

We check the version, roles, MCP configuration and network exposure of your LiteLLM proxy, apply the update and, if required, take over ongoing operation with CVE monitoring.

  • You speak directly with Timo and Robin, who are technically responsible
  • A professional assessment of your situation
  • Concrete next steps for infrastructure, software or AI

The conversation is free and non-binding.

WZ-IT's advice on our Azure migration was technically sound and completely non-binding right from the intro call - we took away a great deal.
Jakob ÖschlbergerInno7 GmbH
Timo and Robin, founders of WZ-IT

How do you run LiteLLM?

WZ-IT's advice on our Azure migration was technically sound and completely non-binding right from the intro call - we took away a great deal.
Jakob ÖschlbergerInno7 GmbH