LiteLLM and Starlette RCE bugs scanned in the wild

Critical Vulnerability 1 source

This weekend, we started seeing requests for /mcp-rest/test/tools/list hitting our honeypots. These requests appear to attempt to exploit CVE-2026-42271 and CVE-2026-48710. Horizon3.ai published a sho

By Yazoul AI · automated

What to do

Review the technical details in this article, check if your organization is affected, and apply the recommended security measures. Monitor official vendor advisories for patches and updates.

What Happened

SANS Internet Storm Center reported over the weekend of November 15-16 that its honeypots began receiving requests for the path /mcp-rest/test/tools/list. The traffic appears designed to probe for two chained vulnerabilities: CVE-2026-42271 in LiteLLM and CVE-2026-48710 in Starlette. The scans follow a public technical writeup by Horizon3.ai, which published exploit details for the LiteLLM flaw. The combination of a public proof-of-concept and observable scanning activity is the classic precursor to broader opportunistic exploitation.

LiteLLM is a widely deployed open-source proxy that normalizes API calls across OpenAI, Anthropic, Azure, Bedrock, and other model providers. Starlette is the ASGI framework underneath FastAPI and countless Python services, including LiteLLM itself. The two CVEs are being probed together because they can be combined into a single intrusion path against the same target.

Why It Matters

AI gateways have quietly become high-value infrastructure. A LiteLLM proxy typically holds provider API keys, model routing rules, usage logs, and sometimes user prompt history. Compromise does not just yield code execution on a container; it can yield the credentials that bill an organization’s entire model spend, plus access to internal model endpoints and any data flowing through them. Because these services are often stood up quickly by ML teams rather than hardened by platform security, they tend to sit outside established patch and monitoring workflows.

The Starlette issue compounds this. If it permits the URL-parsing bypass described in the advisory, an attacker can slip past routing or authentication middleware that assumes normalized paths, reaching internal handlers that were never meant to be internet-facing.

Technical Details

The observed request targets /mcp-rest/test/tools/list, an endpoint tied to LiteLLM’s Model Context Protocol (MCP) integration. That surface is relatively new and, in many deployments, unauthenticated by default. The LiteLLM flaw (CVE-2026-42271) enables remote code execution; the Starlette flaw (CVE-2026-48710) is a URL-parsing bypass that can be used to reach the endpoint through otherwise protective middleware. Horizon3.ai’s publication provides the exploitation blueprint that scanners are now automating.

Indicators to hunt for: inbound requests to /mcp-rest/ paths, especially test/tools/list, from unfamiliar sources; unusual outbound connections from LiteLLM containers; and new or modified processes spawned inside those containers.

Immediate Risk

Risk is critical for any organization exposing a LiteLLM proxy to the internet, particularly one reachable without authentication and running a version predating the vendor fix. Exposure is not limited to large enterprises: the ease of standing up LiteLLM means the vulnerable population is broad and includes smaller teams. Given active scanning and a public PoC, the window between probe and exploit is short.

Priorities: identify all LiteLLM and Starlette deployments, patch both, restrict the /mcp-rest/ endpoints to trusted networks, rotate provider API keys if a proxy was internet-facing, and review logs for the path above. Detection guidance for both issues is available in our LiteLLM RCE advisory and the Starlette URL bypass advisory. Check your exposure now.

Security Insight

This is the Redis and Elasticsearch story repeating itself in a new layer of the stack. Those data stores became mass-compromised because convenience-first defaults put them on the internet without authentication; LiteLLM’s MCP test endpoints carry the same DNA. The non-obvious defensive takeaway: treat AI gateway endpoints as a distinct inventory class with their own exposure review, rather than folding them into generic container scanning. If your asset database cannot answer “which of our services expose /mcp-rest/,” the next wave of scanning will answer it for you.

Further Reading

Share

Never Miss a Critical Alert

CVE advisories, breach reports, and threat intel — delivered daily to your inbox.