“Every enterprise in EMEA is being asked the same question by its board, its regulators, and its customers: ‘How are you governing the AI systems you’re deploying?’”
The EU AI Act, DORA and NIS2 form a regulatory trifecta that demands rigorous governance of AI. The problem: traditional security approaches break down for AI agents, which make autonomous decisions and can be manipulated through prompt injection. Controls have to be enforced architecturally, not asked for in a system prompt.
A defense-in-depth architecture in five layers:
- Network isolation first — agents operate inside constrained boundaries (Private Link, network policies, optional inference proxies) to prevent unauthorized egress.
- Authentication via identity propagation — autonomous agents use short-lived service identities; human-delegated agents carry the human’s identity to preserve audit trails and RBAC.
- Authorization that doesn’t trust prompts — platform-enforced RBAC, masking and row-level security guard data regardless of agent manipulation.
- MCP is a new attack surface — tool poisoning, injection via tool returns and confused-deputy problems need explicit validation and secure credentials.
- Purpose-built beats generic — in regulated environments, agents that embed security natively outperform generic frameworks with security bolted on.
This is the backbone of how I design sovereign AI for regulated clients — and why my own SMCP adds authentication and encryption to MCP.
This is a condensed version. Read the full article on kevinkeller.org →