“As autonomy increases, trust cannot be implied.
It must be proven.”
— Pablo Fourez, CDO Mastercard · March 5, 2026
“Safe scaling demands IBA first.”
— Grok, xAI · Session 10 · March 21, 2026
“External verification > internal promises.”
— Grok, xAI · Session 12 · March 24, 2026
The Seven Ungoverned Sins of Agentic AI
I
Thou shalt not deploy an agent without a signed intent certificate declaring its authorized scope before execution begins.
II
Thou shalt not mistake authentication for authorization. Knowing who the agent is does not declare what it is permitted to do.
III
Thou shalt not govern at the model layer. A model that judges its own actions is not governed. It is self-policing.
IV
Thou shalt not use monitoring as prevention. An audit log of the breach is not a gate before it.
V
Thou shalt not allow a natural language command to expand a cryptographic intent boundary. The certificate does not read prompts.
VI
Thou shalt not deploy a hyperagent — one that modifies its own improvement process — without declaring the boundaries of self-modification before iteration zero.
VII
Thou shalt not launch an autonomous agent at planetary scale — satellite, robot, biological — without a cryptographic gate between declared human intent and execution.
The Catechism of Intent
Questions asked before the agent acts. Answers encoded before execution begins.
Q: What is the authorization gap?
The space between what a human declared and what an agent executed. It exists in every agentic deployment that lacks a cryptographic intent boundary. It is not a bug. It is an architectural absence.
Q: What is an intent certificate?
A cryptographically signed declaration of authorized scope, issued before the agent acts. It specifies the authorized target, permitted operations, declared boundaries, and temporal validity. It is the commandment the agent cannot disobey.
Q: Can an intent certificate be jailbroken?
No. A prompt injection cannot reframe a cryptographic boundary. An instruction chain cannot override a signed certificate. The gate does not read instructions. It validates actions against the certificate. The certificate is immutable.
Q: Who holds authority to sign the certificate?
The human who bears accountability for the agent’s actions. The certificate forces the declaration of authority before execution begins. It does not assume the governance structure is correct. It demands it be declared explicitly.
Q: What happens to an action outside the certificate?
It fails at the gate. Before execution. Before the network call. Before the SIEM entry. Before the audit log. Before the breach. The unauthorized action never occurs.
Q: Is this new?
The architecture is new. The principle is ancient. Humanity has always bound power to declared intent — in constitutions, commandments, contracts, and law. IBA Intent Bound Authorization is the version that runs in under 1.4ms.
✦
GO IN PEACE. GOVERN YOUR AGENTS.
The silicon layer declared itself intelligent this week.
The authorization layer that governs it already exists.