Critical Vulnerability in MCP Server Platform Exposes 3,000 Servers and Thousands of API Keys

A crucial vulnerability in Smithery.ai, a preferred registry for Mannequin Context Protocol (MCP) servers. This challenge might have allowed attackers to steal from over 3,000 AI servers and take API keys from 1000’s of customers throughout many companies.

MCP powers AI apps by linking them to exterior instruments and knowledge, like native filesystems or distant databases. Servers are available native or distant flavors, with distant ones typically self-hosted or totally managed by suppliers.

In accordance with GitGuardian, Smithery.ai’s hybrid mannequin simplifies deployment by internet hosting user-submitted servers on its infrastructure, constructed from GitHub repos into Docker pictures. However this comfort amplified the stakes: a single breach might ripple throughout a whole ecosystem of AI instruments.

Exploiting a Easy Configuration Vulnerability

The flaw stemmed from lax controls in Smithery’s construct course of. Customers submit a smithery.yaml file specifying the Docker construct context through dockerBuildPath. Legit setups level contained in the repo, however the system didn’t validate inputs, enabling path traversal assaults.

By setting dockerBuildPath to “..”, attackers might reference the builder machine’s residence listing exterior the repo, exposing delicate recordsdata to a malicious Dockerfile.

In testing, GitGuardian crafted a repo named “take a look at” with a rigged yaml and Dockerfile. The latter used curl to exfiltrate the listing tree to an attacker-controlled website, revealing recordsdata like .docker/config.json.

This file held an overprivileged fly.io authentication token, meant for Docker registry entry however granting broader machine API privileges.

Fly.io powers Smithery’s internet hosting with virtualized containers, and the token unlocked a company with 3,243 apps, principally MCP servers, plus service infrastructure.

With the token, attackers might question apps, execute code on machines (confirming root entry through “id” command), and even sniff community site visitors.

Compromised Server key

Capturing HTTP requests to a compromised server uncovered client-sent API keys, like a Courageous key in question params. Scaled up, this might harvest secrets and techniques from 1000’s of purchasers connecting to companies through MCP servers, in keeping with GitGuardian.

The incident highlights supply-chain perils in centralized AI internet hosting. MCP servers typically depend on static API keys moderately than OAuth, easing assaults however complicating privilege limits.

Echoing breaches like Salesloft’s OAuth abuse, it reveals how one flaw permits lateral motion throughout trusts.

Smithery mounted the traversal on June 15, 2025, after disclosure on June 13, rotating keys and tightening builds. As AI ecosystems develop, such platforms should prioritize isolation to protect builders from ecosystem-wide threats.

Observe us on Google Information, LinkedIn, and X for each day cybersecurity updates. Contact us to characteristic your tales.