Glow Labs, a security research group at Glow Security, reported on November 3 that AI coding agents often install deprecated MCP servers. MCP, the Model Context Protocol, is a standard that lets AI agents connect to outside tools. The researchers reviewed coding session transcripts and saw the pattern across Claude Code, Cursor, Codex CLI, and models from Anthropic, OpenAI, and xAI. In 98% of cases, they say, the agent chose the outdated package itself, apparently from its training data. Glow found four high-severity flaws in these archived servers, including a remote code execution bug in the Puppeteer browser server that prompt injection (hidden instructions planted in content an AI reads) can trigger.
The maintainers say the server was archived in May 2025 and will not be patched. Deprecated packages have been downloaded more than 17.3 million times from npm since July 2025. No exploitation has been seen so far, according to the report. Teams that let agents install tools should inventory which MCP servers are actually running. Glow also points to its own product category as one remedy.
