A safety playbook for enterprise AI agents: Agent Development Lifecycle, unique identities, least privilege, MCP gateway, sandboxing, and continuous audit.
Shift
Agents are probabilistic and adaptive. Design evaluation-first, not code-first.
Lifecycle
Plan, code, test, deploy, operate, and monitor with security in every phase.
Threats
Prompt injection, privilege climb, data leak, attack amplification, and policy drift.
Controls
Constrain agency, permission with RBAC, sandbox execution, and keep a human on irreversible acts.
Gateway
Put an MCP gateway between agents and tools for auth, DLP, quotas, and a kill-switch.
Sample
A support agent can look up an order and draft a refund. Exporting a customer list to a personal inbox is denied.
Framework
Identity, protect, defend, and comply. Unique credentials. No shared service accounts.