NewApp Monitoring is liveSign up free
AgentbeamBeta
AI agent vulnerability · LiteLLM

CVE-2026-114612: LiteLLM: Privilege Escalation to Proxy Admin via Cross-Domain Reuse of the Salt Key

highCVSS 4.0 7.7 · 3.1 9.9GHSA-7hp6-4w63-5g45

LiteLLM proxy uses one salt key both to encrypt stored secrets and to mint session tokens. An authenticated `internal_user` can abuse that to forge a `proxy_admin` token and then run commands on the host through the MCP stdio endpoint. Fixed in 1.100.4, 1.101.3, 1.102.2, 1.103.1 and 1.104.0rc2, depending on your release line.

Published Oct 27, 2026 · Analysis Oct 27, 2026

What an attacker can do

The flaw is a cross-domain key reuse. The proxy encrypts values with the same key it uses for session tokens. A low-privilege internal user asks for a new API key and puts a forged admin credential in a metadata field as the secret value. The proxy encrypts that payload and returns it. When the user then sends the returned value as a bearer token, the proxy decrypts it, trusts the forged admin identity and grants full administrative access (CWE-345, confused deputy).

From proxy admin, the attacker can reach the MCP stdio endpoint. That endpoint starts local commands, so the result is arbitrary command execution on the proxy host. The CVSS 3.1 vector rates scope as changed, because the proxy host and everything it can reach fall under attacker control.

For teams, the attacker only needs a valid internal-user key. An LLM gateway holds upstream provider keys for every model it fronts, so one escalation exposes all of them, plus spend controls and team data. Any employee or service account with an internal-user key is a starting point.

Who is affected

Per the advisory, versions later than 1.91.0 are exploitable in the default configuration. Versions 1.87.0 through 1.90.x are exploitable only when EXPERIMENTAL_UI_LOGIN=true was set. The package ranges in the advisory data start at 1.91.0 and end before each patched release: below 1.100.4, 1.101.3 on the 1.101 line, 1.102.2 on the 1.102 line, 1.103.1 on the 1.103 line, and 1.104.0rc2 on the 1.104 line.

Check the version:

pip show litellm

For Docker deployments, check the image tag. Also check whether EXPERIMENTAL_UI_LOGIN is set in the proxy environment.

litellm

Ecosystem
PyPI
Affected
>= 1.91.0, < 1.100.4
Fixed in
1.100.4
Check
pip show litellm

litellm

Ecosystem
PyPI
Affected
>= 1.101.0, < 1.101.3
Fixed in
1.101.3
Check
pip show litellm

litellm

Ecosystem
PyPI
Affected
>= 1.102.0, < 1.102.2
Fixed in
1.102.2
Check
pip show litellm

litellm

Ecosystem
PyPI
Affected
>= 1.103.0, < 1.103.1
Fixed in
1.103.1
Check
pip show litellm

litellm

Ecosystem
PyPI
Affected
>= 1.104.0rc1, < 1.104.0rc2
Fixed in
1.104.0rc2
Check
pip show litellm

How to fix it

Upgrade to the patched release for your line: 1.100.4, 1.101.3, 1.102.2, 1.103.1 or 1.104.0rc2.

pip install --upgrade "litellm[proxy]==1.103.1"

Replace the version with the patched one for your line. If you cannot upgrade immediately, set EXPERIMENTAL_UI_LOGIN=false. That disables the vulnerable authentication path. It also breaks CLI SSO and Claude Code gateway login, so plan for that.

After patching, rotate the salt key and the upstream provider keys if you suspect abuse.

How to tell if you were hit

Review key-generation requests from internal_user accounts, especially /key/generate calls with unusual or long metadata values, or values that look like tokens. Then look for a bearer token that does not match any key in your database but is accepted with admin rights.

Check proxy admin activity from non-admin users: new keys, teams or model entries, and calls to the MCP stdio endpoint. On the host, look for unexpected child processes of the proxy and outbound connections after those calls.

Original advisory

From the GitHub Advisory Database (CC BY 4.0).

Summary

An authenticated user with internal_user privileges can escalate to proxy_admin and achieve remote code execution. The proxy uses one encryption key for two purposes: sealing secrets at rest and minting session tokens. An attacker exploits this by requesting a new API key with a crafted metadata field containing a forged admin credential as the "secret" value. The proxy encrypts and returns this payload. When this value is then presented as a bearer token, the proxy decrypts it, trusts the forged admin identity, and grants full administrative access, including the ability to execute arbitrary commands via the MCP stdio endpoint.

Affected versions

Versions later than 1.91.0 are exploitable in the default configuration. Versions 1.87.0 through 1.90.x are exploitable only if EXPERIMENTAL_UI_LOGIN=true was explicitly set.

Mitigation

Upgrade to a patched version as soon as possible.

If you cannot upgrade immediately, set EXPERIMENTAL_UI_LOGIN=false. This disables the vulnerable authentication path but also breaks CLI SSO and Claude Code gateway login.

Impact

An internal user can escalate to proxy admin and execute arbitrary commands on the host.

Discovery Credit: Hoa X. Nguyen (OPSWAT Unit 515)

Credited to Haruna38.

Questions

Do I need to be an admin to exploit this LiteLLM flaw?

No. An authenticated user with internal_user privileges is enough. The attacker uses that account to get the proxy to mint a token carrying a forged admin identity.

Does setting EXPERIMENTAL_UI_LOGIN=false protect me?

It is the advisory's workaround. It disables the vulnerable authentication path, but it also breaks CLI SSO and Claude Code gateway login. Upgrading is the real fix.

Which LiteLLM versions are fixed?

1.100.4, 1.101.3, 1.102.2, 1.103.1 and 1.104.0rc2. Pick the one on your release line. The advisory credits Hoa X. Nguyen of OPSWAT Unit 515 for the report.

References

How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. The analysis is written by an AI model from the sources listed above and checked by automated rules before it is published. Check versions against the vendor's advisory before you act. Methodology

Coverage on the Beam blog

More LiteLLM vulnerabilities

All LiteLLM entries