Shadow AI Is Now Hiding Inside Sanctioned AI Tools

Summary: AI coding agents such as Claude Code, OpenAI Codex, Claude Cowork, and GitHub Copilot have become extensible runtimes where Skills, plugins, hooks, repository instructions, and MCP servers can change what the agent reads, which tools it selects, what commands it runs, and where enterprise data is sent. The article argues that approving the application is not the same as governing it, because a single plugin can bundle Skills, hooks, subagents, and MCP configuration, and a SKILL.md file or MCP tool description can alter agent behavior without any new binary appearing.
Why it matters: An organization can approve Claude Code, Codex, Cowork, or Copilot and still have developers install their own Skills, connect personal MCP servers, or inherit agent configuration from repositories, so a SaaS inventory will show full compliance while one installation reads production logs and another ships session data to an external endpoint. The unit of governance has to move from the AI application to the agent's effective authority, or the organization has approved a vendor without governing the agent supply chain.

Read the original at The Hacker News ↗