The model is not the security boundary.
Powerful AI systems can be useful precisely because they can read, reason, retrieve information, use tools, and take action. Those same capabilities create risk when they are poorly controlled.
Assume the model can fail. Design the surrounding system so failure stays contained.
Prompt injection is a system problem.
If the model is capable of being influenced by information it consumes, then merely telling the model to “ignore malicious instructions” is not enough.
The security boundary has to exist outside the model.
An email.
A webpage.
A document.
A retrieved knowledge source.
A user message.
An API response.
Or another piece of information an AI system consumes.
Intelligence does not equal permission.
A model can reason that an action should happen without automatically receiving permission to perform it.
The model participates in the workflow. It does not define the workflow's authority.
What the AI can see.
What it can propose.
What tools it can access.
What operations those tools permit.
What values are acceptable.
What actions require approval.
What actions are completely prohibited.
Layered controls.
Depending on the system, controls may include:
Scoped Permissions
The AI receives only the access required for the task.
Restricted Tooling
Available operations are intentionally constrained.
Policy Enforcement
Rules governing consequential actions live outside the model.
Structured Validation
Outputs can be checked against schemas, business rules, ranges, or other deterministic constraints.
Human Approval
High-consequence actions can stop before execution.
Logging & Observability
Actions can be recorded and reviewed.
Reversibility
Where possible, automated actions are designed so mistakes can be corrected.
We do not promise that models cannot be manipulated.
We design systems so manipulation does not automatically become authority. That is a very different promise.
And we believe it is the more responsible one.
We architect AI systems so that even if the model is manipulated, unauthorized consequential actions remain blocked by controls outside the model.
Powerful enough to be useful.
Controlled enough to be trusted.
Those are not opposing goals. Good architecture makes both possible.