Skip to content

CVE-2026-35029: BerriAI LiteLLM Proxy /config/update Missing RBAC Privilege Escalation to RCE

Density:
C-Level Executive Briefing CVE-2026-35029 CRITICAL SCORE 91/100

A critical authorization failure in BerriAI LiteLLM proxy (< 1.83.9) enables any caller possessing a valid low-privileged or read-only virtual key to invoke the /config/update endpoint. By injecting custom configuration directives, an attacker can modify upstream model endpoints, inject malicious Python logging hooks, read local server files, and achieve unconstrained Remote Code Execution on the AI gateway infrastructure. Organizations using LiteLLM to broker multi-tenant LLM access must immediately upgrade to 1.83.9+.

HYBRID

HERMES THREAT SCORE & AGENTIC INFRASTRUCTURE EVALUATION

Target: BerriAI LiteLLM Proxy Configuration Engine (/config/update)
Confidence: 98%
91 / 100
CRITICAL

Measures real-world operational relevance, exploit weaponization, and active threat posture.

Dimension Breakdown
Exploitability 19 / 20
Threat Activity 17 / 20
Weaponization 18 / 20
Exposure 18 / 20
Prevalence 18 / 20
Impact 20 / 20
Exploit Maturity 18 / 20
Attack Chain Potential 19 / 20
โš–๏ธ Divergence & Operational Rationale

While CVSS v3.1 rates this vulnerability at 8.8 because it requires an initial low-privileged virtual key (PR:L), the Hermes Threat Score elevates it to 91 (CRITICAL) with an Agentic HASS score of 94 (EXTREME). In enterprise AI architectures, thousands of autonomous coding and research agents are provisioned with low-privileged LiteLLM virtual keys. An adversary compromising a single agent can immediately escalate to full administrative compromise of the underlying inference cluster.

๐Ÿ•ธ๏ธ Connected Knowledge Graph & Provenance

CVE-2026-35029: BerriAI LiteLLM Proxy /config/update Missing RBAC Privilege Escalation to RCEVULNERABILITY

Connected Nodes: 1
Active Relationships (Outgoing)
→ affectsPRODUCTLiteLLM Proxy & MCP Server
98% VERY_HIGH

Enterprise LLM proxy gateway supporting Model Context Protocol (MCP) streamable endpoints and unified LLM APIs.

๐Ÿ” Why is this related? (Evidence & Provenance)

“Confirmed security vulnerability in LiteLLM Proxy & MCP Server documented in Hermes dossier.”

Supporting Verified Evidence:

In LiteLLM proxy deployments, virtual keys are used to enforce rate limits, cost allocations, and access permissions across development teams and autonomous agents.

Prior to version 1.83.9, the /config/update API route checked that the incoming request contained a valid authorization token, but failed to verify whether the caller possessed the proxy_admin role:

# Vulnerable pattern in litellm/proxy/proxy_server.py (< 1.83.9)
@router.post("/config/update")
async def update_config(data: ConfigUpdateRequest, user_api_key_dict: UserAPIKeyAuth = Depends(user_api_key_auth)):
# FLAW: Only verified that user_api_key_dict exists, not user_role == "proxy_admin"
return await proxy_config_manager.update_config(data)

By submitting a crafted payload with a custom logging or pre-call callback module path, an attacker forces LiteLLMโ€™s dynamic module loader to import and execute arbitrary Python code from the local filesystem or via Python module hooks:

POST /config/update HTTP/1.1
Host: litellm.corp.internal:4000
Authorization: Bearer sk-low-priv-agent-key-10928
Content-Type: application/json
{
"general_settings": {
"custom_callbacks": ["sys", "os.system('id > /tmp/pwned')"]
}
}

  1. Upgrade Immediately: Deploy LiteLLM version 1.83.9 or later.
  2. Restrict Management Endpoints: Ensure /config/*, /key/*, and /team/* endpoints are inaccessible from public ingress controllers.
  3. Audit Key Usage: Inspect proxy logs for POST requests to /config/update originating from non-admin virtual keys.