AI Security reference architecture
AI agent security
Scoped permissions, approvals and monitoring for autonomous and semi-autonomous agents.
Design principles
- Agents act on behalf of an identity
- High-impact actions need a human
- Sandbox code execution
- Baseline and alert on behavior
Components
| Component | Role |
|---|---|
| Agent registry | Approved agents and owners |
| Scoped credentials | Short-lived, task-bound access |
| Approval workflow | Human review for risky actions |
| Behavior monitoring | Anomaly detection |
More AI Security architectures
AI security architecture
A security layer between applications and models covering prompts, data, agents, MCP and policy.
View architectureMCP security
Identity, authorization, inspection and audit for agents calling tools through Model Context Protocol servers.
View architectureLLM governance
Inventory, risk tiers, approved providers and review gates for enterprise use of language models.
View architectureAdapt this to your environment
We'll help you apply this pattern to your constraints and existing estate.

