# The Security Stack Playbook

> No agent ever sees the whole vault. Bootstrap key, vertical vaults, machine identity, monthly audit.

*OP3 · Module M7 — How I Run My Day · The CEO AI Playbook by Stephen Forte*

Canonical: https://academy.buildclub.com/m7/op3-security-stack

## Thesis

No agent ever sees the whole vault.

## Key Takeaways

- No agent ever sees the whole vault.
- 90% — of AI agents are over-permissioned — they hold rights they never use.
- 3 of 10 — OWASP Agentic Top 10 risks tied directly to identity, secrets, and over-scoped access (LLM06, LLM07, LLM08).
- 30 days — is the rotation cadence that catches drift before an auditor or attacker does. Calendar-locked, not best-effort.
- Inventory every secret an agent currently has access to. Most CEOs find at least one that surprises them.

## The Playbook

1. **Bootstrap key stays offline** — On a hardware key or a sealed envelope. Touched only at machine provisioning. Never on a laptop hard drive.
2. **Vertical vaults, not horizontal** — One vault per service domain (email, cloud, billing, code). Never one vault for everything.
3. **Machine identity per agent** — Each agent authenticates as itself, with scoped permissions. No shared service accounts. Identity is the audit unit.
4. **Separate accounts for separate blast radii** — Production data is not in the same cloud account as experiments. A blown experiment never reaches a customer record.
5. **Monthly audit, calendar-locked** — Last Friday of the month. Review identities, rotate keys, check the audit log. The discipline survives because the calendar enforces it.

## Key Numbers

- **90%** — of AI agents are over-permissioned — they hold rights they never use. (OWASP Agentic Security Initiative)
- **3 of 10** — OWASP Agentic Top 10 risks tied directly to identity, secrets, and over-scoped access (LLM06, LLM07, LLM08). (OWASP LLM/Agentic Top 10)
- **30 days** — is the rotation cadence that catches drift before an auditor or attacker does. Calendar-locked, not best-effort. (Operator discipline)

## What to Do Monday

- Inventory every secret an agent currently has access to. Most CEOs find at least one that surprises them.
- Pick a secret manager (Infisical, 1Password, Vault). Move every plaintext secret out of code and out of env files this month.
- Split your vaults vertically. Email creds in one. Cloud creds in another. Refuse to merge them.
- Calendar-lock the monthly rotation. Last Friday of the month. Treat it like board prep, not chore time.

## FAQ

**What is the core idea of The Security Stack Playbook?**

No agent ever sees the whole vault.

**What does the data say a CEO should pay attention to?**

90% of AI agents are over-permissioned — they hold rights they never use. 3 of 10 OWASP Agentic Top 10 risks tied directly to identity, secrets, and over-scoped access (LLM06, LLM07, LLM08).

**What should a CEO do Monday morning after reading The Security Stack Playbook?**

Start here: Inventory every secret an agent currently has access to. Most CEOs find at least one that surprises them; Pick a secret manager (Infisical, 1Password, Vault). Move every plaintext secret out of code and out of env files this month; Split your vaults vertically. Email creds in one. Cloud creds in another. Refuse to merge them.

**What are the steps in The Security Stack Playbook?**

1) Bootstrap key stays offline; 2) Vertical vaults, not horizontal; 3) Machine identity per agent; 4) Separate accounts for separate blast radii; 5) Monthly audit, calendar-locked.

**Where do the claims in The Security Stack Playbook come from?**

The playbook cites Infisical machine identity model; OWASP Agentic Security Initiative; OWASP Agentic Security Initiative — over-permissioned agents; Infisical — machine identity for agents.

## Sources

- [Infisical machine identity model](https://infisical.com)
- OWASP Agentic Security Initiative
- OWASP Agentic Security Initiative — over-permissioned agents
- Infisical — machine identity for agents

---

Work with BuildClub: https://buildclub.com/engagements