Google Cloud is treating AI agents as first-class identities.

What Google is recommending

Google Cloud's MCP security documentation says agent-only operation is vulnerable to prompt injection, insecure tool chaining and naive error handling. Its production authentication guidance recommends a separate agent or workload identity instead of reusing a human identity.

Google also recommends granting only the minimum permissions required and using controls that prevent read-write MCP access to important resources where appropriate. Its newer MCP server documentation combines identity, IAM policies and Model Armor around tool calls.

Why identity changes the security model

If an agent runs under an administrator's credentials, the model inherits an administrator-sized blast radius. A prompt injection can become an authorization problem because the agent already holds enough authority to perform the harmful action.

Dedicated identity makes policy enforceable. The system can distinguish one coding agent from another, limit resources, revoke access, separate read and write duties, and attribute actions to the workload that actually performed them.

Where RaksHex fits

RaksHex should treat identity as the first input to every consequential decision. The useful chain is agent identity → delegated authority → semantic action and arguments → policy → credential mediation → execution → Action Ledger.

RaksHex should not claim that it replaces cloud IAM. Cloud IAM decides what a principal can access inside that cloud. RaksHex's opportunity is to add action-level authorization and delegated-agent context across tools and providers before a credential or tool call is released.

Primary sources

• Google Cloud — MCP AI security and safety
• Google Cloud — Authenticate to MCP servers
• Google Cloud — MCP servers overview