The Best Guardrail Is What the Model Can't See
We said guardrails belong in the tool, not the prompt. Connecting an AI to a real home sharpened that: the strongest limit is scope.
Earlier we wrote that guardrails belong in the tool, not the prompt. Our founder has now connected an AI assistant to his home, and the build sharpened the point.
1. Scope beats refusal
The first version refused to unlock doors. The connected version can't see the locks at all. A refusal can be argued with; something that isn't in scope can't be reached. Before the first connection, about sixty items were removed from what the model can see, including the equipment that keeps the house online.
2. Default-deny, without the maintenance
The home network now blocks any device it has never seen, pending approval. The usual approach, an allow-list of hardware addresses, fails because nobody maintains it. Here, the known devices are learned once and anything new requires a decision. Safe by default, cheap to run.
3. Check the device, not the theory
On day one, the guard flagged an unknown "iPhone." The model's first explanation, a Wi-Fi extender rewriting addresses, was confident and wrong; the owner has no extender. Querying the devices directly showed digital photo frames that generate a new hardware address on every restart. The rule was fixed from the evidence, not the theory.
4. Access you can revoke
The assistant signs in with the homeowner's own account and holds a key that can be revoked at any time. It can be trusted because it can be turned off.
Never assume, verify. That's the discipline behind Fi, the platform our products run on.
The full write-up: Describe It, Part Two.