Security researchers at GitGuardian uncovered a critical path traversal vulnerability in Smithery.ai, a Model Context Protocol (MCP) server hosting platform. The flaw, stemming from improper validation of the `dockerBuildPath` parameter, allowed attackers to access arbitrary filesystem locations on Smithery’s build infrastructure. Exploiting this, attackers could exfiltrate sensitive files, including Docker authentication credentials stored in `.docker/config.json`, which were overprivileged and granted access to Smithery’s Docker registry and fly.io’s machines API.The compromise enabled arbitrary code execution on over 3,000 hosted MCP servers, exposing API keys, authentication tokens, and network traffic from thousands of clients. The centralized nature of Smithery.ai amplified the risk, creating a supply chain attack vector where a single breach could cascade across dependent organizations. While no evidence of malicious exploitation was found before patching, the incident highlighted severe risks in AI infrastructure security, particularly around static, long-term API keys and centralized hosting models. Smithery.ai remediated the issue within 48 hours of disclosure, but the vulnerability posed a high-risk scenario for large-scale data exposure and infrastructure takeover.