Start with actors and trust boundaries
The first difference is who interacts with the system and what they can reach. An internal copilot serves authenticated employees and works with internal content. Customer-facing AI serves anonymous users and accepts open input. Each trust boundary calls for its own threat model.
| Dimension | Internal copilot | Customer-facing AI |
|---|---|---|
| Who interacts | Authenticated employees | Untrusted, anonymous users |
| Dominant threats | Oversharing, indirect prompt injection through internal content, and excessive permissions | Adversarial input, abuse, cross-tenant leakage, and brand and legal exposure |
Identity and permission inheritance
An internal copilot usually acts with employee permissions, so it inherits user credentials and, with them, whatever access that employee has. Customer-facing AI should instead rely on scoped application identities.
Data exposure and tool action surfaces
Data and retrieval
Internal copilots retrieve from enterprise-wide content, which makes oversharing and indirect prompt injection through internal content the central concerns. Customer-facing AI draws on tenant-bound and public-facing data, where cross-tenant leakage matters most.
Tool and agent actions
Internal copilots take actions on behalf of an employee. Customer-facing AI exposes actions to the public, so consequential tool calls warrant review.
Threat model comparison at a glance
Five dimensions separate the two threat models.
| Dimension | Internal copilot | Customer-facing AI |
|---|---|---|
| Actors and trust boundaries | Authenticated employees and internal content | Anonymous users and open input |
| Identity and permissions | Inherited user credentials | Scoped application identities |
| Data and retrieval | Enterprise-wide content | Tenant-bound and public-facing data |
| Tool and agent actions | Actions taken on behalf of an employee | Actions exposed to the public |
| Monitoring and audit | Accountability for individual users | Abuse detection at scale |
Prioritizing controls and runtime governance
Internal copilots
Prioritize in this order:
- Clean up source permissions and sensitive-content exposure.
- Constrain tool scopes.
- Treat retrieved content as untrusted.
- Log data access and tool calls with user attribution.
Customer-facing AI
Prioritize:
- Input and output controls.
- Tenant isolation.
- Narrowly scoped service identities.
- Rate limiting and abuse monitoring.
- Review of consequential tool calls.
Neither order is fixed, and your own data and architecture should decide it.
Design-time controls will not catch everything, because model behavior depends on runtime context. A runtime governance layer is worth considering where it addresses a specific threat you have identified: policy enforcement on prompts, retrieved content, and tool calls; least-privilege agent permissions; approval workflows for sensitive actions; and audit logging that supports oversight. Trussed AI provides runtime governance and security for enterprise AI agents in these areas, but the same requirements apply to any approach you choose.
Criteria for evaluating a runtime governance platform
- Can it apply separate policies to internal and customer-facing deployments, with clear scoping and versioning?
- Does it enforce policy at runtime on prompts, retrieved content, and tool calls, rather than only at design time or after the fact?
- How does it handle agent identity, including delegated user credentials versus scoped service identities?
- Can it restrict agents to least-privilege tool access and require approval for sensitive actions?
- Does it log data-access and tool-call decisions, and can those logs reach your SIEM and oversight workflows?
- For customer-facing use, how does it support tenant isolation and abuse monitoring, and what independent evidence exists?