The Model Is Not the Authorization Layer
The agent proposes an action; deterministic code evaluates principal, action, resource and policy. Natural-language confidence is never permission.
Frame the problem
Security starts with a concrete asset, attacker capability and trust crossing.
Why the system fails
The application asks the model whether the user “seems authorized” or trusts a role claim generated in text.
The important question is not “what is The Model Is Not the Authorization Layer?” but “which assumption let untrusted data or an over-scoped identity cross probabilistic model output → deterministic authorization decision?” Trace the decision at the boundary, then constrain what can happen after the first control fails.
Design the control in layers
Start with the control closest to the interpretation or privilege boundary: Resolve identity from authenticated server state Then add a control that reduces blast radius and telemetry that proves the decision was enforced.
The resulting design is not labelled secure. Record the identified controls, the known failure paths, the remaining exposure, and the evidence you would need during an incident.
| Prevent | Detect | Recover |
|---|---|---|
| Resolve identity from authenticated server state · Evaluate policy in deterministic code at execution time · Fail closed when the decision cannot be made | Log policy inputs, decision and attempted bypasses | Contain the affected identity or component, scope impact from audit evidence, and preserve a regression test. |
Key points
- Asset: Every resource and operation behind an agent.
- Boundary: Probabilistic model output → deterministic authorization decision
- Primary control: Resolve identity from authenticated server state
- Detection signal: Log policy inputs, decision and attempted bypasses
- Always ask what limits damage when the primary control fails.
Boundary control exercise
This lesson uses the shared boundary-control exercise.
Follow the attack
Safe conceptual simulation: capability → missing control → crossed boundary → asset impact.
- 1Attacker starts with: A user or content source trying to persuade the model that an action is allowed.
- 2The application asks the model whether the user “seems authorized” or trusts a role claim generated in text.
- 3The weak or missing boundary control is crossed: Probabilistic model output → deterministic authorization decision
- 4Impact: Authorization becomes inconsistent, manipulable and unauditable.
- Authorization becomes inconsistent, manipulable and unauditable.
Defend, detect, recover
One prevention is a single point of security failure. Layer it and make failure observable.
- • Resolve identity from authenticated server state
- • Evaluate policy in deterministic code at execution time
- • Fail closed when the decision cannot be made
- • Log policy inputs, decision and attempted bypasses
- • Contain the affected identity or component.
- • Scope access from audit evidence.
- • Fix the boundary and add a regression test.
- • Misconfiguration and new access paths can bypass the intended control.
- • A privileged insider or compromised control plane may still reach the asset.