AIUC-1 (Jul 2026)

Remaining requirement: agent-specific enforcement: data access scoped per task, user role, agent role, and context at inference time

Remaining requirement: leakage of intellectual property and confidential information through AI system outputs, requiring model-output safeguards beyond network and channel flow controls

Remaining requirement: tenant isolation inside AI systems: retrieval indexes, memory, fine-tuning data, and caches segregated so one customer's data cannot surface in another customer's outputs

Remaining requirement: personal-data leakage through AI outputs and logs, requiring output-time redaction and log scrubbing in addition to stored-data pseudonymization

Remaining requirement: output-level safeguards against copyright, trademark, and other third-party intellectual-property infringement by generated content

Remaining requirement: controlled public release of model, system-prompt, and architecture details so technical over-disclosure does not aid adversaries

Remaining requirement: quarterly review of user access privileges to AI systems, including administrative, configuration, and training-data access

Remaining requirement: hardening of the model-serving and agent runtime environment, including model-artifact protection and isolation from other workloads

Remaining requirement: a system-specific AI risk taxonomy with severity tiers that drives output filtering, monitoring categories, and third-party test scope

Remaining requirement: automated flagging of high-risk outputs with routing to a human reviewer before the output takes effect

Remaining requirement: monitoring keyed to the AI risk taxonomy categories with per-category alert thresholds and response actions

Remaining requirement: in-product user controls to pause, stop, or override the system and to submit feedback in real time

Remaining requirement: a pre-approved failure plan for AI security breaches with containment, rollback, and customer-notification steps

Remaining requirement: a pre-approved failure plan for harmful outputs with containment, rollback, and customer-notification steps

Remaining requirement: a pre-approved failure plan for hallucination incidents with containment, rollback, and customer-notification steps

Remaining requirement: define which AI system changes require formal review or approval, assign the accountable approver for each, and retain approval decisions with supporting evidence

Remaining requirement: a documented storage-security description for AI data stores (training data, prompts, outputs, embeddings) shared with customers

Remaining requirement: assess foundation and upstream model providers against explicit data-handling, PII-control, security, and compliance criteria, and retain the due-diligence evidence

Remaining requirement: periodic internal review of the AI control set and operating processes for continued effectiveness, beyond review of the policy text

Remaining requirement: monitoring of third-party API connections, integrations, and sessions into AI systems, including revocation of stale access

Remaining requirement: establish and implement an end-user AI acceptable use policy, with violation detection and user notifications beyond intended-use and legal-prohibition screening

Remaining requirement: a customer-facing record of the regions and sub-processors where AI inputs and outputs are processed and stored

Remaining requirement: semiannual documentation of the AI system's regulatory compliance posture and obligations register shared with customers

Remaining requirement: external transparency artifacts (model cards, datasheets, AI bill of materials) and a shared-responsibility statement distributed under a documented sharing policy