# On-Premise AI Security Checklist

## Architecture and data boundary

- [ ] Orchestration location is documented.
- [ ] Model-serving location is documented.
- [ ] Retrieval, memory, and logs have documented storage locations.
- [ ] Network egress rules are enforced.
- [ ] Sensitive workflows can operate without external model calls.

## Model policy and guardrails

- [ ] Approved local and frontier models are listed by workflow and data class.
- [ ] Prohibited data is blocked before model calls.
- [ ] Region and provider policy is enforced.
- [ ] Tool requests are validated independently from model output.
- [ ] Guardrail decisions are logged.

## Identity, permissions, and approval

- [ ] Workflows use dedicated service identities.
- [ ] Read and write scopes follow least privilege.
- [ ] Delegated authority expires.
- [ ] Sensitive actions require named approval.
- [ ] Prohibited actions are technically blocked.

## Evidence and operations

- [ ] Tool calls, retries, approvals, and changed records are recorded.
- [ ] Logs are access-controlled and protected from modification.
- [ ] Retention and deletion policies are documented.
- [ ] Abnormal activity and repeated policy denials are monitored.
- [ ] Recovery is tested for model, integration, and internal-system failures.

Source: Dialogo, reviewed 2026-08-15. https://www.dialogo.chat/resources/on-premise-ai-security-checklist/