BadHost flaw hits Starlette, exposes MCP servers
A path-bypass vulnerability in Starlette (CVE-2026-48710) allows attackers to breach AI tools via malformed HTTP Host headers. The trivial exploit affects vLLM, LiteLLM, and most FastAPI-based MCP servers, exposing stored credentials and third-party account access.
The AI ecosystem's deep reliance on FastAPI and Starlette means a single web framework bug just became an industry-wide infrastructure crisis. Because MCP servers inherently store credentials to external tools, this bypass turns agent harnesses into massive attack vectors.
- –A single character injected into the HTTP Host header bypasses path-based authorization in Starlette versions prior to 1.0.1
- –The vulnerability affects major inference engines including vLLM, LiteLLM, and Text Generation Inference
- –MCP servers are uniquely impacted as they typically hold API keys and database credentials to execute agent tool calls
- –The flaw is trivial to exploit against systems not protected by a properly configured firewall
- –Discovered by X41 D-Sec, the bug carries a 7.0 severity score that researchers argue severely understates its downstream impact
DISCOVERED
15h ago
2026-05-28
PUBLISHED
18h ago
2026-05-28
RELEVANCE
AUTHOR
Hrethric