Skip to content

Local or delivered entropy, an argument about trust boundaries

Entropy delivered over a network and entropy generated locally differ on four axes: availability, verifiability, sovereignty and attack surface. A shared service is a reasonable engineering choice in some settings, but it moves the trust boundary — and moving a trust boundary is an architectural decision, not a deployment detail.

Argued on the merits, including where the delivered model is defensible. An argument that concedes nothing persuades nobody.

Where a shared service is reasonable

If entropy is needed occasionally, on platforms with no hardware source, and the consuming system already trusts a provider with far more sensitive material, then drawing randomness from a service is not obviously wrong. The provider may well operate better sources than the consumer could.

Saying so plainly matters, because the case against is about trust boundaries rather than about the model being invalid.

Availability

Delivered entropy has a network dependency. If key generation blocks on a remote call, then a partition, an outage or a rate limit becomes a cryptographic failure — and the moment you most need fresh keys is often the moment after a disruption, when everything reconnects at once.

Verifiability

When entropy arrives over a link, what you can verify is that bits arrived over an authenticated channel. Whether they were unpredictable at their origin, whether they were unique to you, and whether the source was healthy at that moment are all statements you take on the provider's word.

A local source with health reporting turns those into measurements. That is the whole difference, and it is why local generation is the stronger position rather than merely the preferred one.

Sovereignty

For defence, government and regulated infrastructure, the origin of key material is a jurisdictional question as much as a technical one. Entropy crossing an organisational or national boundary is often simply not permitted, regardless of how good it is.

Attack surface

A remote source adds a channel, an authentication mechanism, a provider and a set of intermediaries. Each is defensible individually. Together they are a larger surface than a component inside the device, and they sit between the source and the thing that needs to trust it.

The decision, stated properly

Not "which is better" but "where should the trust boundary sit". If the answer is inside the device, entropy has to be generated there — which is a question about form factor as much as anything: where the source physically goes.