Security checklist
On-Premise AI Security Checklist for Agentic Workflows
A buyer-oriented checklist for local LLMs, local infrastructure, network boundaries, model guardrails, permissions, approvals, and run evidence.
Deployment boundary
Document where orchestration, models, retrieval, memory, logs, and connected tools run. In a fully local deployment, the architecture can be designed so prompts, retrieved context, workflow state, and run evidence remain inside the company network.
Model policy
Define which workflows may use local models, which may use approved frontier models, and which data classes may never cross the network boundary. Hybrid deployments should enforce policy before any model call, not rely on user memory.
Frontier-model guardrails
Where approved frontier models are used, a guardrail engine should classify data, remove or block prohibited context, enforce model and region policy, validate tool requests, and record the decision. Sensitive workflows can remain fully local.
Identity and permissions
- Use separate service identities by workflow or risk domain.
- Grant least-privilege read and write scopes.
- Separate tool access from model access.
- Expire delegated authority.
- Require named approval for sensitive actions.
- Record the acting identity, mandate, and approver.
Evidence and operations
- Log tool calls, returned status, retries, approvals, and changed records.
- Protect logs from modification and restrict access.
- Define retention and deletion rules.
- Monitor abnormal volume, new tools, policy denials, and repeated failures.
- Test recovery when a model, integration, or internal system is unavailable.
Security review outcome
The review should end with a documented architecture, data-flow diagram, permission matrix, model policy, approval matrix, logging plan, incident owner, and test evidence. “On premise” is a location; security depends on the controls implemented around it.
Start with one workflow and prove the boundary.
Map the event, systems, permissions, approvals, exceptions, and evidence before granting access. The result is a pilot scope your operations and technical teams can evaluate together.
Book a workflow mapping