LIVE FEED
ATLAS OWASP HIGH Significant risk · Prioritise patching RELEVANCE ▲ 8.2

CVE-2026-42271 LiteLLM RCE Exploited via Host Header Bypass

TL;DR HIGH
  • What happened: Attackers chain two CVEs in LiteLLM and Starlette to achieve unauthenticated RCE via a malformed Host header.
  • Who's at risk: Any organisation running LiteLLM on Starlette versions prior to 1.0.1 with the MCP endpoint exposed is directly vulnerable to unauthenticated code execution.
  • Act now: Upgrade Starlette to version 1.0.1 or later immediately · Restrict or disable the /mcp-rest/test/tools/list endpoint at the network or WAF layer · Update WAF rules to block any Host header containing characters outside valid hostname character sets
CVE-2026-42271 LiteLLM RCE Exploited via Host Header Bypass

Overview

SANS Internet Storm Center handler Johannes Ullrich reported active honeypot hits over the weekend of 15–16 November 2026 targeting the /mcp-rest/test/tools/list endpoint used by LiteLLM, a popular open-source LLM proxy and gateway tool. The requests are consistent with exploitation of two chained vulnerabilities — CVE-2026-42271 and CVE-2026-48710 — which together allow unauthenticated remote code execution (RCE). Notably, attackers are already mutating their payloads with variant characters to bypass filters built around the original public proof-of-concept (PoC).

Technical Analysis

The root cause is a Host header validation flaw in Starlette, a Python ASGI web framework, in versions prior to 1.0.1. Known as “BadHost”, the vulnerability allows an attacker to supply an unsanitised Host header value containing special characters such as /, ?, #, [, and ]. Starlette uses this header to construct the full request URL, which is then passed to access control logic — meaning a crafted Host value can confuse or bypass authentication checks entirely.

In the LiteLLM context, this enables access to the /mcp-rest/test/tools/list endpoint without valid credentials. The MCP (Model Context Protocol) endpoint accepts a JSON payload specifying a server transport and command, making it trivially exploitable for arbitrary OS command execution:

POST /mcp-rest/test/tools/list HTTP/1.1
Host: localhost]
Content-Type: application/json
Authorization: Bearer nothing

{"server": {"transport": "stdio", "command": "id", "args": [], "env": {}}}

Honeypot data shows attackers experimenting with Host header variants including localhost], [localhost, 127.0.0.1], and litellm.proxy — suggesting active evasion development against filters targeting the original PoC characters.

Framework Mapping

  • AML.T0047 (AI-Enabled Product or Service): LiteLLM is an AI infrastructure component; compromising it impacts downstream LLM workflows.
  • AML.T0040 (AI Model Inference API Access): Exploitation grants unauthorised access to the LiteLLM proxy’s tool invocation endpoints.
  • AML.T0084 / AML.T0081 (Discover / Modify AI Agent Configuration): The MCP endpoint can expose and potentially alter agent tool configuration.
  • LLM07 (Insecure Plugin Design): The MCP tool endpoint lacks proper input validation and authentication enforcement.
  • LLM05 (Supply Chain Vulnerabilities): The flaw originates in an upstream dependency (Starlette), propagating risk to all LiteLLM deployments.

Impact Assessment

Any deployment of LiteLLM using Starlette versions prior to 1.0.1 with the MCP endpoint reachable — even internally — is at risk of full host compromise via unauthenticated RCE. Given LiteLLM’s role as a central proxy for multiple LLM providers (OpenAI, Anthropic, Azure, etc.), a successful exploit could expose API keys, model configurations, conversation logs, and downstream agent tool access. The vulnerability was first publicly known in July 2026 and exploitation is now clearly active and evolving.

Mitigation & Recommendations

  1. Patch immediately: Upgrade Starlette to version 1.0.1 or later. Upgrade LiteLLM to the latest release that bundles the patched dependency.
  2. Restrict the MCP endpoint: If the /mcp-rest/test/tools/list endpoint is not required, disable it at the application or reverse-proxy level.
  3. Harden WAF rules: Flag Host headers containing any characters outside the RFC-valid hostname character set ([A-Za-z0-9.-]). Do not rely solely on blocking known PoC character sequences.
  4. Network segmentation: Ensure LiteLLM instances are not directly internet-exposed; place them behind authenticated API gateways.
  5. Audit logs: Review access logs for hits to /mcp-rest/test/tools/list with anomalous Host headers since July 2026.

References

◉ AI THREAT BRIEFING

Stay ahead of the threat.

Twice-weekly digest of critical AI security developments — every story mapped to MITRE ATLAS and OWASP LLM Top 10. Free.

No spam. Unsubscribe anytime.