SANS Internet Storm Center reports that honeypots started seeing POST /mcp-rest/test/tools/list over the weekend. Those requests look like attempts at CVE-2026-42271 in LiteLLM, the AI gateway proxy, chained with CVE-2026-48710 in Starlette. The sample body asks for a stdio transport with command: id. Somebody is checking whether the box will run commands for them.
The LiteLLM bug, per NVD and the project’s own advisory: from 1.74.2 up to before 1.83.7, two endpoints used to preview an MCP server before saving it (/mcp-rest/test/connection and /mcp-rest/test/tools/list) accepted a full server config in the request body. That includes command, args and env for the stdio transport. The endpoint connects, which means it spawns your command as a subprocess on the proxy host, with the proxy’s privileges. The only gate was a valid proxy API key. No role check. A low-privilege internal-user key was enough. NVD lists a CVSS 4.0 score of 8.7, high.
Read that again. An endpoint whose job is “test this config” executed the config. The save endpoint already required admin; the test endpoints didn’t. That’s the oldest shape of bug there is: two doors to the same room, and only one has a lock.
The chain is what makes the scans interesting. According to SANS, Starlette before 1.0.1 doesn’t validate the Host header (“BadHost”), builds the full URL from it, and then makes access control decisions on that URL. Stuff /, ? or # into the Host header and the access check gets confused. Horizon3.ai wrote up how the two combine into unauthenticated remote code execution. The scans SANS saw use Authorization: Bearer nothing and Host headers containing [, which looks like someone fiddling with the public PoC to dodge filters. SANS also notes this isn’t new; exploitation was reported back in July.
What to do:
- Upgrade LiteLLM to 1.83.7 or later. The fix makes both test endpoints require
PROXY_ADMIN. - Upgrade Starlette to 1.0.1 or later. I’m taking that version from SANS; check what your pinned dependency tree actually resolves to rather than trusting the top-level requirement.
- If you can’t upgrade today, the advisory says to block both POST paths at your reverse proxy or gateway.
- Filter on the class of bug, not the PoC. SANS’s advice is right: reject any Host header containing characters that can’t appear in a hostname, or better, allow only the hostnames you serve. A deny-list of the exact characters in last July’s exploit is how you lose to the second week’s exploit.
- Ask why your LLM gateway is reachable from the internet at all. It holds provider keys and can now spawn processes. That is a lot of privilege for something that mostly forwards JSON.
Opinion: the MCP stdio transport is “run this program” with a protocol wrapper on it. Any endpoint that takes that config from a request body has to be treated like ssh with a web UI. I’d love to say this is a new class of problem. It’s system() from a form field, which we all agreed in 1998 was a bad idea, and then the AI plumbing crowd reinvented it with better documentation. For the habits that blunt this sort of thing, see the sandboxing hub and the kernel and security hub. Same story as yesterday’s MongoDB failover bug: reachable before anyone has really checked who you are.
