Skip to content
Blog

Field notes on agent security

Working notes on governing AI coding agents at the endpoint: shadow MCP, endpoint governance, and verifiable custody. No fluff.

Topics:architectureauditbeaconbyokendpointgovernanceinventorymcpshadow-mcptrust
  • 6 min read

    Shadow MCP: the inventory problem nobody owns

    MCP made every developer an integrator and every hand-edited editor config a piece of unmanaged infrastructure. No CPE, no package record, no gateway log — and no tool in your stack that models it.

    shadow-mcpmcpinventory
    Brian McManusRead →
  • 5 min read

    Your AI gateway can't see the agents that matter

    A gateway governs exactly one thing: traffic that agrees to route through it. The local Claude Code, the hand-added MCP server, the tool call that never leaves the laptop — governance has to live where the call originates.

    beaconendpointgovernance
    Brian McManusRead →
  • 6 min read

    Cut us off: the BYOK kill-switch drill

    A security vendor asking for your trust should hand you the scissors. By default we hold an envelope key we delete on request (a crypto-shred); in every BYOK engagement you revoke your own key and watch our access die in your own CloudTrail.

    byoktrustaudit
    Brian McManusRead →