Key facts
- SANS Internet Storm Center honeypots detected requests this weekend for /mcp-rest/test/tools/list that appear to target CVE-2026-42271 and CVE-2026-48710.
- Horizon3.ai published a write-up explaining that chaining CVE-2026-42271 and CVE-2026-48710 achieves remote code execution without authentication.
- Starlette releases older than 1.0.1 lack Host header validation, a flaw known as BadHost.
- A published proof of concept confuses access control logic by placing /, ? and # characters in the Host header.
SANS Internet Storm Center honeypots recorded requests this weekend for the /mcp-rest/test/tools/list endpoint that seem connected to CVE-2026-42271 and CVE-2026-48710. Honeypots are decoy systems that attract and detect malicious traffic. The scanning source remains unattributed.
Horizon3.ai released a write-up stating that combining CVE-2026-42271 and CVE-2026-48710 produces remote code execution without any authentication required. CVE is a Common Vulnerabilities and Exposures identifier. Remote code execution means an attacker can run commands on the server. Johannes B. Ullrich, Ph.D., Dean of Research at SANS.edu, authored the SANS Internet Storm Center diary published Nov. 16, 2026.
Starlette is a Python web framework. Versions of Starlette older than 1.0.1 do not validate the Host header, a flaw known as BadHost, which means 1.0.1 or newer performs the validation. Starlette constructs a full URL using the unvalidated Host header and relies on that URL for access control choices. LiteLLM relies on Starlette to handle access control, leaving the /mcp-rest/test/tools/list endpoint open despite its usual authentication requirement, which verifies a user’s identity before granting access.
An existing proof of concept posted online confuses access control logic when /,? and # characters appear in the Host header. Scanning observed by the honeypots uses Host header entries like localhost], [localhost, 127.0.0.1 and litellm.proxy, including bracket characters that might slip past filters built for known exploits. People behind the attacks adjust known exploits to slip past filters and locate additional targets.
A single captured request carries User-Agent Python/3.14 aiohttp/3.14.5, which identifies the client software making the request, and contains a JSON body that executes the id command, signaling a remote code execution attempt. The id command is a standard system command that prints user identity information. Attacks surfaced in July, meaning this is not a freshly uncovered flaw, and the exploitation predates the current scanning surge. Johannes B. Ullrich said, “This isn’t a new vulnerability, and exploitation has been reported back in July.”
The request captured by the honeypots carries an Authorization header set to Bearer nothing. Its body size appears as Content-Length 74. The diary shows this raw HTTP request as one example of the scanning traffic.
It welcomes every media type and names gzip, deflate and zstd as acceptable encodings. The JSON body designates stdio as the transport method, id as the command to run, an empty args array and an empty env object, attempting to launch a system command through stdio. JSON is JavaScript Object Notation, a data format. stdio is standard input/output, a way for programs to receive input and send output.
Protective measures cover patching, shrinking the attack surface, checking hostnames before building URLs, and WAF rules that reject disallowed hostname characters or permit only certain hostnames. WAF is a web application firewall that filters and monitors HTTP traffic. Ullrich said firewall rules ought to aim at the flaw rather than individual exploits. He added that checking a hostname before assembling a URL is simpler than trying to validate and restrict URLs inside middleware, which is software that connects different applications or services, when that middleware has to accept many URL types.
Ullrich said shrinking the attack surface is likely the key defensive measure after patching during the present wave of new vulnerabilities. He said developers ought to check data parts before merging them into complicated structures. The diary presents these defenses as mitigation guidance.

Join the discussion