Australian cloud for business software — built and run in Sydney since 2002

NetLeverage
NetLeverage

Encryption keys for customer-held backups: who holds them and what losing or copying one means

Technical article
Backup & Recovery

About this article: This article is reviewed against the cited sources. This is source research, not lab validation. Check the stated versions and your own requirements before making changes.

If a backup copy delivered to you is encrypted, the key or passphrase is part of the backup. If every key that opens the copy is lost, the copy cannot be read, by you, by your provider or by anyone else. So the question of who holds the keys is a decision about recovery, not only about security.

Two things are easy to confuse. Losing the only key is total loss, but losing one key is not total loss if another valid key survives. And giving a second person a copy of the key improves recovery while making it easier for the data to be read by the wrong person, because anyone holding a working key can open the copy alone.

Our recommendation, which the sources support in parts but do not state as one rule, is that you should always be able to retrieve at least one working key under your own control, without depending on the provider being available. A provider that holds a working key can help with restores, but it can also read the data, so decide that deliberately and write it down.

Why it matters

A lost key does not cause a slow restore or an expensive one. If it was the last working key, it causes no restore. Backup tool documentation says so plainly. The restic documentation, for example, states: "Please note that knowledge of your password is required to access the repository. Losing your password means that your data is irrecoverably lost." (restic, Preparing a new repository.) It repeats the warning: "Remembering your password is important!" That warning describes a repository with a single way in. The same documentation describes repositories with several passwords, which changes what losing one means. Restic is one tool among many, used here because its documentation is public and clear, not because any provider uses it.

The same property can be turned against you. If an attacker obtains a working key, the encryption stops protecting the copy. The restic reference says an adversary with a leaked key could "Decrypt existing and future backup data", and that "a password change will not prevent this" in the setup described, because changing the master key "can currently only be done using the copy command, which moves the data into a new repository with a new master key, or by making a completely new repository and new backup." (restic, References.) The master key is the single key that actually encrypts the data; each password only unlocks a stored, encrypted copy of it.

So a key must be kept available enough to be recovered and secret enough not to be stolen. Those two aims pull in opposite directions, which is why custody needs an explicit decision.

What the key standards say

NIST's key management recommendation (Special Publication 800-57 Part 1 Revision 5, May 2020) is written for government systems, but its reasoning about availability and recovery applies to anyone. Two points are relevant.

First, NIST says key information may need to be readily available for as long as data is protected by the key, and that "A common method for providing this protection is to make one or more copies of the key information and store them in separate locations." (NIST SP 800-57 Part 1 Rev. 5, section 6.2.2.1.) It describes backup storage as "an independent, secure storage media" that provides "a source for key recovery". In our reading, a key written down only on the machine it protects, or stored next to the backup it unlocks, does not meet that description.

Second, NIST treats backing up a key as a decision, not a default. It says "Not all keys need to be backed up", and where operations can continue by issuing new keys, "it may be preferable not to save the keying material in order to lessen the possibility of a compromise". Its table lists a symmetric data encryption key as OK to back up (permissible, not required). For a backup copy the choice usually points one way. A key that protects old backups cannot be replaced by a new one, because the old data stays encrypted under the old key. In our reading, a backup key is the clear case for keeping a protected copy. Note the trade-off NIST itself names: every saved copy raises the possibility of compromise.

NIST also warns about retention. Key information retained after the end of its active period "should be placed in archive storage". If your retention policy keeps backups for seven years, the keys needed to read them must last as long. Keep that in your key record.

Redundant custody is not split knowledge

Two arrangements are often described as sharing a key, but they do opposite jobs.

Redundant custody means several holders each have something that opens the copy on its own: two sealed envelopes each containing the full passphrase, or two separate passwords that both unlock the same data. It protects against losing a key, because any one surviving holder can recover. It does not protect against a holder misusing or leaking theirs, because any one holder is enough. Adding holders increases availability and increases exposure together.

Split knowledge is a different technique. The NIST glossary defines it as "A process by which a cryptographic key is split into n key shares, each of which provides no knowledge of the key." If knowledge of k shares is required to construct the key, then, in NIST's words, "knowledge of any k – 1 key shares provides no information about the key other than, possibly, its length." (NIST CSRC glossary, split knowledge.) No single holder can open the copy; a quorum of k holders must act together. The cost runs the other way: if more than n minus k shares are lost, nobody can rebuild the key. It needs a recombination process that someone has tested.

So two sealed full-key copies are not a simpler form of split knowledge. They are redundant custody, and either holder can decrypt alone. The restic pages we read describe passwords, not key shares; if you want a quorum rule, treat it as a separately designed and tested process, and do not assume your backup tool provides it.

Custody options

The table is our framework, built on the sources above. None of the options is right for everyone.

Arrangement Who can open the copy alone What it gives you What can go wrong
Provider holds the only working key The provider Easy restores through the provider You cannot read your own copy if the provider is unavailable, or if the relationship ends. The copy is not independent.
One key, one customer holder That one person or system Full independence If that key is lost and no other valid key exists, the data is gone. The provider cannot help with restores unless it is given a key.
Redundant custody: several full copies of one secret, or several passwords on the same data Any single holder Losing one copy or one holder does not lose the data Any one holder can leak or misuse it. Removing a password later does not undo what a holder already copied (see below).
Split knowledge: k of n shares Only k holders acting together One holder alone cannot open the copy or hand it over Losing more than n minus k shares loses the key. The recombination process must be tested.
Key lodged with a trusted third party The third party, on the terms agreed Resists one person's absence or error The third party must itself be trusted and reachable, and you rely on its terms.

A key held by your provider is a provider ability to read as well as to restore. If you want exclusive custody, no one else holds a working key. If you give the provider a working key so that it can help with restores, record that as provider-held, not customer-exclusive.

Several passwords on the same data

Some backup tools support more than one key for the same data. The restic documentation says "A repository can have several different passwords, with a key file for each", and "The key command allows you to set multiple access keys or passwords per repository." It describes list, add, remove and passwd sub-commands. That makes redundant custody practical: two people can each hold a different password for the same data.

What the sources also say matters for how you rely on it. The key files "contain all data that is needed to derive the repository's master encryption and message authentication keys from a user's password", and the reference notes that several passwords exist so that "the password can be changed without having to re-encrypt all data." In other words, every password wraps the same master key. Three consequences follow, and the first two are our reading of those statements:

  • Removing a password's key file from the repository stops that password opening the repository as it stands. It does not change the master key. A person who already holds a copy of that key file and its password, or a copy of the data and a derived master key, still holds what they hold.
  • Changing or removing a password therefore does not protect data against someone who already had access. The restic reference states that "With the current key management design, it is impossible to securely revoke a leaked key without re-encrypting the whole repository."
  • Protection for the future needs a new master key: move the data into a new repository under a password only the right people hold. Anything that was already copied out under the old key stays readable by whoever copied it. Treat it as exposed.

Adding a key needs an existing one. In restic's documented example, the add command first prompts "enter password for repository" and then "enter password for new key" (restic, Encryption). So the person adding a key must already hold a working one, and the new password is best typed by the person who is meant to own it, at their own prompt, rather than sent to someone else. If a provider or contractor runs the step for you, the new password has been known to them, and you cannot call it exclusive. Not every tool supports several keys. Ask whether yours does.

What the sources say happens when a key is deleted

Cloud key services illustrate the point at larger scale. AWS describes deleting a key in its Key Management Service as "destructive and potentially dangerous" and says: "After a KMS key is deleted, you can no longer decrypt the data that was encrypted under that KMS key, which means that data becomes unrecoverable." (AWS, Delete an AWS KMS key.) AWS adds that "you cannot recover a deleted KMS key."

AWS builds in a delay as a safeguard. It requires "a waiting period of 7 – 30 days", with a default of 30, during which the key "cannot be used in any cryptographic operations" but deletion can be cancelled. It suggests disabling a key instead of deleting it when unsure.

One more point from the same page matters for customer-held copies: "If you close your AWS account, your KMS keys become inaccessible". If a key lives inside a provider account, losing access to that account may make the key unreachable. We cite AWS only as an example of how cloud key services behave. It does not tell you what any given backup provider does; the questions below are how you find out.

Questions to put to your provider

Ask for the answers in writing.

  • Is the delivered copy encrypted, and who generates the key?
  • Where is the key held, and can you or your staff use it without our involvement?
  • Who, other than us, holds a working key or password for the copy, including any held by the provider?
  • Can the same data be opened by more than one key, and who is authorised to add or remove one?
  • If we lose our key, what can you do? If you lose yours, what happens?
  • How is a key rotated or replaced, and does that affect copies already delivered? If a key may have been exposed, can the data be moved under a new master key?
  • How long are keys retained, compared with how long copies are retained?
  • What happens to keys you hold if our agreement ends?

Read-only checks

You can check what keys exist for a repository that uses restic without changing anything. The following is from restic's documentation, shown for illustration. It was not run for this article. Linux shell:

restic -r /srv/restic-repo key list

The tool asks for a password and lists the keys with an identifier, user, host and creation time; an asterisk marks the one in use. Use your own repository path. The number of lines tells you how many keys are registered in the repository, which you can compare with your key record. It does not show who still holds the password for each one. Treat this as a check that the key you hold still opens the repository, not as a full restore test. For other tools, use their own documentation.

Without a command line, use a key record. Write down, for each encrypted copy:

  1. Which backup set and tool it belongs to.
  2. How many keys open it, and who holds each, including any held by the provider.
  3. Where each key is stored, and what would be needed to retrieve it.
  4. Who else can retrieve it if the first person is unavailable, and whether each holder can open the copy alone or only with others.
  5. When it was last confirmed to work.
  6. How long it must be kept (as long as the longest retained backup that depends on it).

A worked example (illustrative)

The following is invented. A small accounting practice receives encrypted nightly backups into its own storage. The passphrase was set up by its former IT contractor and is stored in the contractor's password manager. The practice owner writes the key record and finds that the only copy of the passphrase is outside the practice's control. The contractor, who holds the passphrase, runs the key listing at the owner's request, and it shows one key. If the contractor's password manager is lost, nothing else opens the data.

The owner asks the contractor, in writing, to run the add-key step while the owner types a new passphrase personally at the second prompt. The repository now has two working keys on the same master key. The practice has gained redundant custody; it has not gained exclusive custody, because the contractor's old passphrase still opens everything. The owner does not give the new passphrase to the provider. If the provider ever needs a working key to help with a restore, the owner will decide that separately and record the provider as a holder.

The owner stores sealed copies of the new passphrase with the practice's solicitor and with a second director, and records that this is redundant custody: either holder can open the copy alone, which is not split knowledge. A colleague test-opens one backup using the new passphrase, and the owner records the date.

Then comes the old passphrase. The contractor knew it, and a password does not change the master key. After the new key is confirmed to work, the owner can ask for the contractor's key file to be removed, which stops that password opening the repository as it stands. But the owner treats the old passphrase, and any copy the contractor made, as exposed. To protect future backups, the practice asks the provider about moving the data into a new repository with a new master key, the route the restic reference describes, and retires the old repository once the new one is verified. Anything copied out earlier stays readable by whoever copied it.

Risks and limits

  • Our recommendation to keep a customer-held key is a judgement, not a NIST requirement. NIST says the final decision to back up keys "should be made based on the application".
  • A copy of a key is a second thing to protect. Every extra holder is another chance of leakage, so record and review them.
  • Removing a password or key file does not revoke access already obtained. Only a new master key protects future data, and it does nothing for copies already taken.
  • Not every backup tool supports several keys or key shares. Where it does not, you may have to choose one custodian or share one key under an agreed procedure.
  • Split knowledge needs a tested recombination process; we have not tested one, and we make no claim that any particular tool supports it.
  • This guide does not assess any provider's key management or compliance obligations. NIST's guidance is written for United States government systems and is used here as a reference, not as a legal requirement in Australia.
  • AWS and restic documentation describe their own products. Other products behave differently, and restic behaviour may differ between versions.

Next step

Write down, for each encrypted backup copy you receive, how many keys open it, where each is held and who can retrieve it, then ask a colleague to confirm they can find it. If you are reviewing a backup service, the Backup and Recovery page is a starting point for that conversation; ask for the scope in writing.

Sources

First accessed 15 November 2026, except the NIST glossary entry, first accessed 16 November 2026; restic pages re-read 16 November 2026.

Learning notes

  • An encrypted backup key is part of the backup. If every key that opens it is lost, the data cannot be read; if one of several is lost, it can still be read.
  • Two sealed copies of the full key are redundant custody: either holder can open the copy alone. Split knowledge, where a quorum of shares is needed, is a different technique.
  • Anyone holding a working key can read the data, including a provider. Do not call a key exclusive if someone else was given a working one.
  • Removing a password does not re-key existing data or take back what a holder already copied; if it may have been exposed, plan a new master key.
  • Write down how many keys exist, who can find each, and check it still works.