PRACTICE · RUNTIME GOVERNANCE

AI Security & Runtime Governance

Security controls at the moment of commitment

Independent deterministic controls for consequential actions by AI-enabled and conventional software systems.

Many AI controls operate at the level of training, prompt design, model evaluation, monitoring, or workflow orchestration. These may be valuable, but they are not substitutes for an independent decision about whether a particular state-changing action may be committed now.

Execution-boundary questions

Is the actor authorized for the proposed operation?

Is the proposed action admissible under the current governing rules?

Does current, attributable evidence support the decision?

Will the attempted transition preserve required invariants?

Separation of responsibilities

AI models and agentic workflows generate proposals; a deterministic enforcement component evaluates authorization, admissibility, evidence, and invariants; the enterprise-controlled system commits authorized actions.

This is summarized as: AI proposes → AGCP evaluates → Enterprise commits.

Engineering and assurance

The architecture must address the freshness of evidence, machine identity, rule-version provenance, treatment of long-running actions, and verification of resulting state. It should not rely on an LLM acting as the final judge of its own consequential actions.