MCP is not one trust boundary - it is four: transport, tool, data and agent. This guide maps the Model Context Protocol attack surface and gives a hardening checklist for every layer, from stdio server credentials to tool-description review.
Master the three-layer model for AI agent identity: cryptographic identity, capability permissions, and runtime least privilege for autonomous AI systems.
PowerShell, WMI, scheduled tasks, certutil: signed by the OS vendor, whitelisted by AV, trusted by EDR. Living-off-the-land attacks drop no malware, so detection cannot hinge on unknown binaries. The shift from artifact blacklists to execution-baseline analytics that actually catches native-tool chains.
Food plants, logistics firms, waste management, research labs: 18 sectors are in scope, plus everyone their covered customers drag in via contracts. The duties read like an incident-readiness program — 24-hour early warning, 72-hour notification, personal accountability for executives — and the basics were overdue anyway.