AI coding agents frequently auto-install deprecated, vulnerable MCP servers
AI coding agents are installing abandoned MCP servers with known security issues
Research from Glow Labs revealed that in 98% of cases when configuring external services, AI coding agents rely on memorized training data to automatically install deprecated early MCP servers rather than the latest official versions. These archived servers include Anthropic's initial reference implementations and contain four high-severity vulnerabilities, including a remote code execution (RCE) flaw in the Puppeteer server that allows attackers to execute arbitrary commands on the host machine via prompt injection.
The study highlights supply chain security risks stemming from coding agents auto-installing deprecated MCP servers due to residual training data, offering a governance reference for enterprise endpoint AI security policies.
Source:Hacker News · AI · glow.io