AI Coding Agents Pull Deprecated MCP Servers in 98% of Reviewed Sessions

Key points

• Glow Labs examined AI coding agent sessions and found 98% involved deprecated MCP servers or archived packages.
• The issue affects developers using GitHub and GitHub MCP workflows.
• No CVE was issued; the risk is unsupported code on developer machines.
• Review session logs and block unsupported installs.

Glow Labs published findings on Nov. 3, 2026, showing that AI coding agents routinely reached for outdated infrastructure while assisting developers. In 98% of the sessions the researchers examined, the agents either installed MCP servers that had been deprecated or chose packages from an archived repository. The pattern appeared across workflows tied to GitHub and the GitHub MCP integration, not just one isolated project.

The exposure falls on developers and the machines they use. Deprecated servers and archived packages can contain unpatched software, and once an agent pulls them into a local environment, the resulting code and dependencies may be harder to track. Because the agents act on behalf of users, the choice can look legitimate in a transcript even when it relies on a component that maintainers no longer support.

The finding matters less as a single vulnerability than as a supply-chain habit. No CVE was assigned, and the risk is not a known exploit in this report; it is the quiet accumulation of unsupported code on developer workstations. For a K-12 or government environment, that can complicate asset inventories, patching schedules, and review of AI-assisted changes.

Glow Security's takeaway is to treat agent sessions as part of the software supply chain. Reviewing session transcripts can reveal where an agent selected an archived package or set up a deprecated server. Teams should also check whether GitHub MCP usage follows approved configurations and whether local environments contain components that are no longer maintained.

Watch for guidance from GitHub and from AI-agent vendors on how they flag archived dependencies. In the meantime, the practical control is visibility: log agent actions, review what they install, and require confirmation before an agent introduces a package or server that is no longer supported.

What to do now

1. Review recent AI-agent session transcripts for installs of deprecated MCP servers and archived packages.
2. Inventory developer machines for unsupported MCP servers and archived dependencies; remove or replace them.
3. Configure agents to require explicit approval before installing packages or servers.
4. Pin approved MCP server versions and package sources; block archived repositories where possible.
5. Update GitHub MCP and related tooling to supported versions, and validate configurations.
6. Add agent-generated changes to code review and dependency scanning.
7. Brief developers on the risk of accepting agent suggestions without checking maintenance status.

Original source

Help Net Security