Definition
Authority leakage occurs when AI systems act outside declared authority boundaries because governance intent is embedded in natural-language prompts or informal engineer interpretation—not in machine-checkable governance contracts.
The organization believes a lending assistant may only recommend within a junior underwriter's limit. The agent, routed through a capable model and a long prompt, produces output that reads like a final decision. Nothing in the runtime enforces the boundary. That gap is leakage.
Why it is hidden
Leakage rarely looks like a security breach. Tasks complete. Outputs look plausible. Moderation passes. The failure is structural: authority is described, not compiled.
- Policy prose lives in SharePoint; the agent never consults it as a constraint object.
- Approval matrices exist on paper; the workflow has no permission predicate.
- Escalation rules are documented; the agent has no routing gate when thresholds are crossed.
Teams discover leakage in audit—not in demo.
Leakage vs. model error
A wrong answer is an accuracy problem. Authority leakage is a governance representation problem. Fixing the prompt may change behavior until the next model swap, platform upgrade, or tool addition. Compilation targets the contract layer—the durable artifact that should survive runtime churn.
What reduces leakage
Governance Compilation treats authority as a first-class constraint: thresholds, roles, escalation paths, and refusal conditions compiled into execution—not paraphrased into instructions.
Pair with governance lineage so auditors can see which authority was invoked, not only what text the model emitted.
Evidence, durability tables, agent-architecture limits, and open research questions:
Open Why Enterprise Governance Doesn't Compile →