Skip to content

Image: isc.sans.edu · rights & removal

Executive Summary

Exploits targeting CVE-2026-42271 and CVE-2026-48710 are being observed via requests to the /mcp-rest/test/tools/list endpoint, which appears to be accessible without authentication. The vulnerability stems from Starlette's failure to validate the "Host" header, allowing input like "localhost]" or "[localhost" to be used when constructing URLs for access control decisions. Attackers are observed experimenting with special characters in the Host header to evade existing filters. This chain of vulnerabilities is known to allow remote code execution without authentication, as detailed by Horizon3.ai. The incident highlights the difficulty middleware has in correctly validating hostname inputs when building complex network access controls.

Facts Only

* Requests were observed hitting /mcp-rest/test/tools/list targeting honeypots.
* These requests appear to attempt exploitation of CVE-2026-42271 and CVE-2026-48710.
* Starlette versions prior to 1.0.1 fail to validate the "Host" header, a vulnerability known as "BadHost."
* The unvalidated Host header is used to build the complete URL for access control decisions.
* Attackers experimented with characters like '/', '?', '#', '[', and '] in the Host header to evade filters.
* Examples of observed Host headers include "localhost]", "[localhost", and "litellm.proxy".
* The exploit involves sending a request with a JSON payload specifying an environment command, such as `{"server": {"transport": "stdio", "command": "id", "args": [], "env": {}}}`.

Full Take

The narrative demonstrates a critical gap between the intended security posture and the reality of operational software, specifically in middleware components like Starlette. The core pattern involves leveraging weaknesses in data validation—specifically hostname parsing—to achieve remote code execution across chained vulnerabilities. The evolution noted where attackers modify payload characters to bypass existing filters illustrates an active, adaptive adversarial landscape where defenses must account for non-standard input combinations rather than just known exploit signatures. The lesson learned emphasizes that defense mechanisms should focus on strict constraint setting—allowing only specific hostnames or strictly validating URL components before assembly—rather than attempting to filter every conceivable malicious string. This suggests a systemic difficulty in applying perimeter security effectively when dealing with dynamic inputs from complex dependencies like those used in AI orchestration layers.
Bridge Questions: How can system architects enforce canonical hostname validation consistently across all middleware, regardless of the upstream framework? What is the cost-benefit analysis for strictly limiting acceptable hostnames versus maintaining flexibility for diverse proxying scenarios? What mechanisms exist to detect the chaining of seemingly disparate vulnerabilities during real-time traffic inspection?

From the original · SANS Internet Storm Center

42271, CVE-2026-48710) 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.
Read the full story at isc.sans.edu

Sentinel unavailable

The automated check of the source article's wording did not complete, so no result is shown.

This looks only at the wording of the original source article, not at this page's AI-written sections. A small local AI model made this estimate. It has not been checked against known human and machine texts, so treat it as provisional. It cannot show who wrote an article.

Scans for LiteLLM/Starlette (CVE-2026 | Huntaegis