CVE-2026-42271: LiteLLM and Starlette exploit scans surge as attackers probe AI gateways
By aminah
·17 Nov 2026
·4 min read

Attackers are actively scanning the internet for LiteLLM servers they can break into through a flaw in the Starlette web framework, according to the SANS Internet Storm Center, which recorded the attempts on its honeypots over the weekend of 14-15 November 2026.
The scans target the /mcp-rest/test/tools/list endpoint and try to chain two vulnerabilities, CVE-2026-42271 and CVE-2026-48710, to run commands on the server without logging in. The SANS diary, written by Johannes Ullrich, says the requests "appear to attempt to exploit CVE-2026-42271 and CVE-2026-48710" and notes that Horizon3.ai has published a short write-up showing how the two flaws can be chained to achieve unauthenticated remote code execution.
LiteLLM is an open-source gateway that lets organisations call different large language model providers through one interface. Many deployments sit in front of model credentials and provider API keys, so a compromised gateway can expose far more than the server itself. SANS warns that the flaw matters most where LiteLLM relies on Starlette for access control.
## How the Starlette flaw opens the door
The root cause is in Starlette, a Python web framework. Versions before 1.0.1 do not validate the Host header, a weakness researchers call "BadHost". Starlette takes the unvalidated header, uses it to build the full URL, and then uses that URL to make access control decisions. An attacker who can confuse that logic can reach endpoints that should require authentication.
The honeypot requests seen by SANS used a POST to /mcp-rest/test/tools/list with a JSON body instructing the MCP stdio transport to run the command "id", a simple reconnaissance probe rather than a weaponised payload. The requests carried malformed Host headers, including "localhost]", "[localhost", "127.0.0.1]" and "litellm.proxy", and a User-Agent identifying a scripted Python client, Python/3.14 with aiohttp/3.14.5.
SANS notes that the published proof-of-concept exploit uses characters such as "/", "?" and "#" in the Host header to confuse the access control logic, and that the attackers observed this weekend appear to be trying different characters, such as brackets, possibly to evade filters built around the known exploit.
## Not a new flaw, but still being exploited
Neither vulnerability is new. CVE-2026-42271 is scored 8.7 and CVE-2026-48710 is scored 6.5, and both are listed in CISA's Known Exploited Vulnerabilities catalogue. SANS points out that exploitation was first reported in July 2026, adding: "But attackers keep exploiting existing vulnerabilities, modifying them to evade filters and find new victims."
That persistence is the warning. Organisations that patched one component but not the other, or that left a LiteLLM proxy exposed to the internet, remain in scope. Reporting on the scans advises upgrading LiteLLM to version 1.83.7 and Starlette to version 1.0.1 or later, and checking logs for POST requests to the MCP test endpoints. Separately, The Hacker News has reported Amazon describing AI agents hunting internet-exposed AI applications, including LiteLLM, to steal provider API keys, which underlines why exposed AI gateways are now a priority target.
## What to check now
SANS draws three lessons from the scans. First, when building firewall or web application firewall rules, focus on the vulnerability rather than on specific exploits: flag every character that should not appear in a hostname, or better still, allow only specific hostnames. Second, software should validate data components, such as a hostname, before assembling them into more complex structures like a URL. Third, reduce the attack surface: patch, and do not expose management or test endpoints to the internet unless they are needed.
Anyone running LiteLLM should confirm the Starlette version in use, apply the fixed releases, and search access logs for requests to /mcp-rest/test/tools/list and the related test endpoints. Unexpected calls to those endpoints should be treated as a possible compromise of the host.
## FAQ
**What are CVE-2026-42271 and CVE-2026-48710?**
CVE-2026-42271 is a flaw in LiteLLM's MCP test endpoints, and CVE-2026-48710 is the Starlette Host header validation flaw. Chained together, they can allow unauthenticated remote code execution on affected LiteLLM deployments.
**Which versions fix the flaws?**
Starlette 1.0.1 or later fixes the Host header validation problem, and LiteLLM 1.83.7 or later is the fixed release for CVE-2026-42271, according to reporting on the SANS findings.
**Is there evidence of successful attacks?**
SANS honeypots recorded exploit attempts, not confirmed breaches. The observed payload ran only the "id" command, which is a reconnaissance step. No follow-on payload or confirmed victim has been reported in the SANS diary.
Sources: SANS Internet Storm Center, "Scans for LiteLLM/Starlette (CVE-2026-42271, CVE-2026-48710)"; Horizon3.ai, "CVE-2026-42271 chained with CVE-2026-48710" (https://horizon3.ai/attack-research/vulnerabilities/cve-2026-42271-chained-with-cve-2026-48710/).
Source
SANS Internet Storm Center — https://isc.sans.edu/diary/rss/33522