The gap between an AI usage policy and its actual enforcement is where compliance nightmares live. On November 9, 2026, Ronan Oliveira of Zihin published a breakdown of how to close that gap, arguing that governance living in a PDF is merely intent, while governance in the runtime is operations. For those of us running autonomous agents, the question isn't just what the agent *should* do, but what it *can* do when nobody is watching at 2 a.m.

Contextual Security Policy Architecture

Oliveira introduces the Contextual Security Policy (CSP), a framework that ties permission decisions to execution context rather than just caller identity. The model relies on five specific axes: Origin (channel/application), Schedule (time windows), Data (field-level access), Behavior (volume/frequency limits), and Custom (domain-specific rules). This isn't a prompt injection or a front-end filter; it’s a hard check executed in the runtime before a tool runs. If the agent tries to cross the line, the call fails. Period.

Inheritance And Auditability

Scalability comes from a strict inheritance model: tenant → agent → user. The organization sets the compliance baseline at the tenant level, and subsequent levels can only tighten restrictions, never loosen them. This ensures that regulations like Brazil's LGPD are defined once and propagated automatically, reducing the surface area for human error. Crucially, Zihin emphasizes that enforcement without evidence is useless. Every policy decision is recorded for session replay, turning audit trails into simple queries rather than weeks-long excavations.

Key Takeaways

  • Runtime enforcement prevents agents from accessing data they shouldn't, regardless of prompt instructions.
  • Inheritance models (tenant → agent → user) allow centralized compliance without per-agent configuration drift.
  • Session replay transforms audit requirements from manual reviews into automated, queryable logs.
  • Zihin positions governance as the necessary counterpart to agent reach, converting compliance risk into an auditable asset.

The Bottom Line

Static policy documents are performative theater for compliance teams; real governance requires code that physically blocks unauthorized agent actions at the runtime layer.