The Strongbox in the Bedroom

A family that installs a strongbox in the bedroom has solved a problem of theft and created a problem of inheritance. The contents are safe from the burglar and inaccessible to the heir; the mechanism that guarantees secrecy also guarantees loss, and the keyholder becomes the single point at which the household's continuity fails. Privacy engineering has arrived at the same architecture. The controls now being mandated are excellent at preventing disclosure and structurally indifferent to the question of who may open the box, and when.

Regulators and procurement offices have begun treating differential privacy, secure multi-party computation, and confidential computing as acceptable substitutes for legal bases and contractual safeguards, converting what are statistical and hardware properties into compliance artefacts. The shift is real, and it is being described in the language of assurance by parties who do not control the assurance.

Two Objections the Enthusiasm Skips

The first is that differential privacy is a calibration, not a guarantee, and its cost is paid in utility by whoever needs the data most. Dwork and Roth set out the framework in their 2014 monograph with unusual candour: the guarantee holds for a specified budget of queries, and the budget is consumed. The monograph is explicit that composition erodes protection across repeated release, which means a programme that publishes one differentially private table a year is defensible and one that publishes weekly is not. Vendors who quote an epsilon without stating the query budget and the composition schedule are selling a number, not a control. Regulators who accept that number as an alternative to a lawful basis are accepting a specification they have not read.

The second objection is architectural and more serious. Confidential computing relocates trust from the counterparty to the silicon vendor, and it does so silently. A workload protected inside a hardware enclave is protected from the operator of the machine and from the tenant sharing it, which is a genuine advance. It is not protected from the manufacturer of the processor, from the firmware signing chain, or from the export-control regime that determines who may purchase the processor. An organisation that outsources its confidentiality to a microarchitecture has not eliminated the trusted third party; it has reduced the number of them to one, and that one is subject to geopolitical process it cannot litigate. Calling this privacy is a category error, and it is being made in boardrooms.

The Precedent: The Safe Deposit and the Sealed Will

The analogous moment is the late-nineteenth-century migration of valuables from the household safe to the bank's vault, accompanied by the simultaneous legal invention of the sealed will. The vault solved custody. It did not solve succession, and the market's answer was not a better vault but a legal instrument that specified who could compel opening, under what evidence, and subject to whose review. The technology and the procedure developed together, and the procedure was the part that survived. Present-day privacy engineering is repeating the first half of that history and skipping the second: the enclave is being procured, and the succession instrument, meaning the key-escrow, audit, and compelled-disclosure framework, is being deferred on the grounds that the hardware made it unnecessary. Hardware does not make governance unnecessary. Hardware makes governance the only remaining control.

What This Does to the Compliance Function

Three consequences follow, and none of them appear in the vendor literature. The first is that the evidence standard changes character: a control that can be attested cryptographically is auditable in a way that a policy is not, and auditable controls attract enforcement faster than discretionary ones, because failure is provable. Organisations deploying these technologies should expect to be held to them precisely, and should not deploy any control they are unwilling to warrant.

The second is that privacy engineering becomes a capital decision rather than a legal one. Enclave capacity, multiparty round counts, and noise budgets are infrastructure line items with depreciation schedules and utilisation targets. The budget conversation moves from the general counsel to the chief technology officer, and with it moves the incentive to understate the residual risk, because the party funding the control is the party measured on its cost.

The third is that a monitoring industry forms around attestation rather than around policy, and it will be run by parties with no legal accountability for the conclusions they certify. Remote attestation reports will be purchased in volume, filed as evidence, and read by no one. That is not an argument against the technology. It is an argument for insisting that the attestation be scoped, time-bound, and tied to a named accountable officer, which is the difference between evidence and ornament.

Primary research, in brief: Industry forecasting has long projected that the majority of data processing would migrate outside centralized facilities, with Gartner's widely cited estimate placing three quarters of data creation and processing at the edge rather than in traditional centralized infrastructure, a distribution that materially weakens any control model premised on a single auditable perimeter.

Directives for Those Buying the Vault

Demand the noise budget, the composition schedule, and the query cap in writing before signing, and treat any vendor unwilling to state them as unsuitable for regulated data. For confidential computing, require disclosure of the attestation chain, the firmware signing authority, and the escrow arrangement, and record the escrow in the same instrument that governs breach notification, so that a hardware failure is a reportable event rather than an outage. Individuals should understand that a privacy label on an application is now a claim about arithmetic, and that the arithmetic is calibrated for the vendor's benefit as much as their own.

Six Months Out

Expect the first enforcement action framed around a misstated privacy budget rather than a missing consent record, because it will be the easiest case to prove and the most instructive to bring. Expect consolidation among attestation providers, and expect at least one procurement standard to be quietly rewritten once a purchaser discovers that its enclave vendor cannot be compelled to produce keys. The strongbox will be installed everywhere. The question of who may open it will remain, as it always did, a matter of law.