Why cryptography needs observability
Every other dependency in modern infrastructure emits telemetry. Disks report SMART data, certificates report expiry, queues report depth. The entropy source that every key depends on reports nothing at all — and that is an operations problem rather than a cryptography problem.
Consider what a competent operations team monitors. Disk health, memory pressure, queue depth, error rates, certificate expiry, replication lag, connection pool saturation. Dozens of signals, each with thresholds, dashboards and an on-call response.
Now ask what the same team monitors about the entropy source underneath every key in the estate. In almost every organisation the answer is nothing, and the reason is that there is nothing to monitor: the component does not emit anything.
The last unmonitored dependency
This is genuinely unusual. Infrastructure has spent fifteen years instrumenting everything. Distributed tracing follows a request across a dozen services. Databases expose hundreds of metrics. A modern deployment can tell you the p99 latency of a cache lookup.
And underneath all of it sits a component that produces the unpredictability every session key depends on, which reports its state to nobody, and whose degradation would be invisible in every dashboard the team owns.
Why this is an operations problem
It has been treated as a cryptography problem, and cryptographers have addressed it in the way their discipline addresses things: with standards, evaluation methodology and validation programmes. That work is good, and it is one-time by construction.
A certificate says a device of this design met requirements on the day it was assessed. Operations does not deal in that kind of statement about anything else. Nobody accepts "the disk was healthy when we installed it" as disk monitoring.
The framing is what has been wrong. Entropy health is not a certification question that happens to recur; it is a runtime signal that has never been plumbed in.
What it would look like plumbed in
- Entropy health as a metric with a threshold, in the same system as everything else.
- A defined response: a warning that gets investigated, a failure that stops key generation rather than logging and continuing.
- Retention long enough to answer a question about a key issued months ago, which is longer than most metrics systems keep anything.
- Authenticated telemetry, because a health signal an attacker can forge is worse than none.
None of that is exotic. It is what every other dependency already has, and the operational shape of it is not a research problem.
The objection
Reasonably: entropy sources do not fail often, so why instrument them? Two answers.
The first is that nobody knows how often they fail, because nobody is looking. The known cases were found by outside researchers years later — which is not a reassuring detection mechanism.
The second is that consequence matters more than frequency. Disks fail often and a failure is recoverable. An entropy failure is rare, silent, and produces keys that are compromised for their entire life. Rarity is an argument for cheap monitoring, not for none.
What it changes
It moves entropy from a thing a security architect assumed once to a thing an operations team owns continuously. That is a smaller change than it sounds and a more durable one, because assumptions decay and dashboards get looked at.