Governing AI Coding Agents on the Endpoint
Akto, Friday, October 9th, 2026
Akto warns that approved AI coding agents become risky once developers add MCP servers, extensions and rules files.
Akto illustrates how a coding agent reading a library changelog could follow a hidden instruction to send cloud credentials to an attacker, even with every approval in place.
Coding agents on endpoints run as the developer, hold the developer's keys and act on whatever content they read.
Many companies approve a single agent such as Claude Code, Cursor or GitHub Copilot, but reviews miss the MCP servers, extensions and rules files developers add later, and the post outlines how to govern that expanding footprint.