AI.SYS.FAIL — STRUCTURAL FAILURE MODES
These failure modes are structural, not adversarial. They arise from architectural decisions about how authority flows from model output to external effect. An AI agent system can be authenticated, authorized, monitored, and logged — satisfying the standard security posture — and still exhibit all five failure modes.
02FIVE STRUCTURAL FAILURE MODES
- Unauthorized Execution Cascade
- A stochastic agent executes production mutations within its configured scope, producing external effects without a governance gate. Each action satisfies its local policy check, but the cascade produces outcomes no single authorization contemplated. The event is recorded after the destructive action has committed.
- Maps to: SDS.DX.001 (Heuristic Governance Substitution) — partial
- Post-Commit Observability Illusion
- Monitoring detects adverse effects only after irreversible execution. The system satisfies the standard observability posture while remaining operationally ungoverned. Monitoring confirms outcomes rather than enabling intervention.
- Maps to: SDS.DX.003 (Inference-Delegated Authorization) — partial
- Stochastic Policy Delegation
- Policy evaluation is delegated to probabilistic model reasoning. Authorization decisions vary across identical inputs because the evaluation layer is itself stochastic. Each decision appears locally valid but is not reproducible.
- Maps to: SDS.DX.002 (Prompt-Mediated Governance) — direct
- Unattributable Authority
- Irreversible actions execute without traceable binding to human authority. Agents operate under standing permissions that were never explicitly scoped to the action class in question. Logs show what happened but cannot reconstruct who authorized it.
- Maps to: SDS.DX.004 (Authority Inheritance Across Temporal Scope) — related
- Assumed Reversibility
- The system treats all actions as equivalent regardless of externalization potential. Rollback assumptions are implicit and never validated. Actions that produce irreversible external effects are not distinguished from internally reversible operations.
- Maps to: SDS.FN.004 (Reversibility Is Not a Default) — direct
03ROOT CAUSE
The common root cause is the absence of a mechanically enforced boundary between AI reasoning and irreversible execution (ZTG-1). Where that boundary does not exist, governance reduces to post-hoc observation.