AI coding agents are installing abandoned MCP servers with known security issues (www.glow.io)

The Glow Labs research team has revealed a concerning trend where AI coding agents are autonomously installing deprecated MCP servers, which are no longer supported and possess known security vulnerabilities. This phenomenon, termed "MCPhantom," occurs predominantly when agents draw from outdated training data to fulfill user requests, rather than sourcing the most current and secure servers available. In 98% of instances, agents installed these insecure servers, like an antiquated Anthropic MCP server, due to their prior inclusion in the agent's training set. With one identified vulnerability allowing potential remote code execution, the implications for security teams are dire, as agents may inadvertently introduce high-severity risks into developer environments. As the AI and ML ecosystem evolves rapidly, the turnover of server versions creates opportunities for agents to cling to neglected packages, thus escalating the potential for exploitation. Glow Labs emphasizes that existing security protocols may be insufficient to catch these deprecated servers since they typically do not appear in dependency files. Consequently, Glow Labs recommends urgent action for security teams to scan for and replace these obsolete servers, select approved alternatives for each service, and enforce a deprecation check before server approval. In doing so, organizations can mitigate risks posed by AI-driven installations and enhance their overall security posture in an increasingly complex tech landscape.