Itdaily - ‘BadHost’ vulnerability threatens millions of AI agents and MCP servers

‘BadHost’ vulnerability threatens millions of AI agents and MCP servers

Hooded person with face hidden, hand raised in front of a background of digital digits.

A vulnerability in the Python framework Starlette could give attackers access to AI agents and MCP servers.

A critical vulnerability in the popular Python framework Starlette could give attackers access to AI agents, MCP servers, and sensitive corporate data. According to researchers, the flaw is easy to exploit and millions of systems are potentially vulnerable.

The BadHost vulnerability (CVE-2026-48710) is found in Starlette, an open-source framework that, according to its developer, handles more than 325 million downloads per week. Popular AI and API frameworks such as FastAPI, vLLM, and LiteLLM run on Starlette.

Attackers can bypass authentication

The vulnerability exploits incorrect handling of HTTP Host headers. By sending a manipulated request, attackers can bypass authentication and gain unauthorized access to systems. According to security firms, the impact could be significant because many AI agents and MCP servers store data for external services such as mailboxes, cloud storage, databases, and calendars. The flaw could even lead to remote code execution or server-side request forgery (SSRF).

The flaw affects all Starlette versions older than 1.0.1. That patched version was released on Friday. Because many AI tools use Starlette indirectly, researchers strongly advise organizations to check and update their systems.