LETTERMCP — Page 3

Enterprises must build tool-level access controls since MCP leaves authorization to operators.

AI agents using MCP can call tools they weren't authorized to use.

Gateways solve the N×M security problem MCP's protocol design left unsolved.

Adversarial prompts and tool abuse expose gaps in Claude's principle-based safety training.

Four boolean hints let hosts read agent risk before tools run, but only if servers tell the truth.

Platform teams now must govern AI agents the way they governed developers.

Governance frameworks built for human workers collapse under the speed and autonomy of AI agents.

OAuth 2.1 replaces static keys with role separation, PKCE, and tokens bound to single servers.

Clarifying host and client roles prevents security gaps in agent governance.

OpenAI's Agents SDK lacks multi-user credential management that enterprises need at scale.

Role clarity determines whether an agent's power stays bounded or spreads through compromise.