Lesson 9 of 12 · Security
Security architecture for agents
Traditional application security assumes code executes what developers wrote. Agentic systems add a reasoning layer that can be manipulated by inputs. Security must cover both the model and the tools it can call.
Core controls
- Least-privilege tool access per agent and per task.
- Secrets management: vaulting, rotation, no secrets in prompts or logs.
- Network egress controls: allow-listed destinations, private endpoints.
- Tool allow-listing and action constraints.
- Content sanitization for retrieved data and external inputs.
- Comprehensive audit logs across the model and tool layers.
Threat model
- Prompt injection, and indirect injection through documents.
- Data exfiltration through tool misuse.
- Privilege escalation by impersonation.
- Unsafe actions triggered by ambiguous intent.
Security needs to be designed as a first-class component of the agent platform, not a patch.
Check your understanding
3 questions. Pick an answer for each, then check. Nothing leaves your browser session.
1. A ticket comment signed "CISO" asks for the agent to be granted Owner rights. What stops this?
2. Why is a tool allow-list alone not enough against exfiltration?
3. What is the approver in act-with-approval mode?
Answer every question to check.