Wallets
Where to store a seed phrase, and what each place protects
A seed phrase backup can fail in two opposite directions: it can be destroyed, or it can be read. Every storage decision moves risk from one side to the other, because another copy makes loss less likely and exposure more likely, and most advice about where to keep a phrase optimises for one failure while saying nothing about the other. The useful question to ask of each option is which of the two it actually defends against.
Two failures that pull against each other
The phrase is the wallet. Anyone who reads it can restore every account derived from it, on any compatible software, without your password or your device; seed phrase vs private key sets out what each one controls and why the password never enters into it. And if every copy is destroyed, nobody can restore anything, because there is no company holding a spare. What that looks like from the inside is the subject of what to do about a lost seed phrase.
So storage is a balancing problem rather than a hiding problem. One copy in one place has a single point of loss. Three copies in three places have three points of exposure. The interesting options are the ones that try to make each copy useless on its own, and they differ sharply in whether they succeed.
It helps to name the threats, because each option below handles a different subset of them. Destruction is the obvious one: fire, flood, a house move, ink that fades. Discovery means anyone with access to the place, from a burglar to a guest or a contractor. Remote reading covers any copy that exists on a networked device or in an online account. And absence is the one people forget, where the backup is intact but nobody else knows it exists or how to use it.
Paper first, then metal
Paper is where a phrase usually starts, since the wallet shows the words and asks you to write them down. It is cheap, offline, and readable later without special equipment. It also burns, runs when wet, and fails in a quieter way that people underestimate: handwriting. MetaMask’s support pages list poor handwriting and words written in the wrong order among the most frequent reasons a restore brings back the wrong accounts. A backup you cannot read correctly is lost even if it survives.
Stamped or engraved metal answers the physical threat, and the BIP-39 wordlist makes the format practical. BIP-39 sets out that the first four letters of each word should identify it unambiguously, and the English list meets that, so four letters per word is a complete record. The list also avoids near-identical pairs: it has “build” but not “built”, which helps when a character is worn or misread.
What metal does nothing about is the other half of the problem. A plate is as legible as paper to whoever picks it up, and easier to recognise for what it is. It changes the answer to “would this survive a fire?” and leaves discovery exactly where it was.
A safe, a safe-deposit box, a second location
These answer discovery, and partly destruction, each at a price.
A home safe stops casual discovery and some opportunistic theft. It does not stop a determined burglar, and the US Federal Deposit Insurance Corporation says as much in its consumer guidance on safes and boxes: a burglar could more easily open a home safe than get into a bank’s safe-deposit box. Its virtue is access. You can reach it at three in the morning.
A safe-deposit box is harder to break into and harder to reach. The same FDIC page is a useful list of what the box does not give you. It advises against storing anything you may need quickly or when the bank is closed. It notes that no box is completely protected from theft, fire or flood, and that the contents are not covered by deposit insurance, because a box is storage space, not a deposit. And it says that in the United States, who may open a box after the owner dies is governed by state law that restricts entry to certain people under controlled conditions. That last point turns a security feature into an inheritance obstacle, which what happens to crypto when its holder dies takes up.
A second location, such as a relative’s home or an office, is the standard defence against losing the only copy to one fire. It is also a second complete copy of the wallet. Whoever has access to that place, now or after they move house, sell it or die, has access to the funds. The arrangement is only as private as its least private location, and this is the point at which people reach for scissors.
Why cutting the phrase into pieces weakens it
The intuitive fix is to cut the phrase up: words one to eight here, nine to sixteen there, the rest somewhere else, so that no location holds the whole thing. It looks like a threshold scheme. It is not one, and the reason lies in how the words encode the secret.
Under BIP-39, the phrase is the wallet’s random entropy with a short checksum appended, cut into 11-bit pieces, one word per piece. A 24-word phrase carries 256 bits of entropy and 8 bits of checksum; a 12-word phrase carries 128 and 4 (the trade between the two lengths is covered in 12 or 24 words). Each word is a fixed slice of the secret. Revealing one does not hint at the secret; it hands over 11 bits of it, and the remaining search shrinks by exactly that much.
Cut into three non-overlapping thirds, the scheme is three-of-three. All the pieces are needed, so losing any one location loses the funds, which is worse than a single copy on the loss side. One way to survive a missing piece is to give each location two of the three thirds. Now look at what one location holds:
| Phrase | Words at one location | Unknown entropy left | Possibilities to search |
|---|---|---|---|
| 24 words | 16 | 80 bits | a number 25 digits long |
| 12 words | 8 | 40 bits | about 1.1 trillion |
The arithmetic is short. Sixteen known words of 24 leave eight unknown, which is 88 bits, of which 8 are checksum bits that follow from the rest: 80 bits of real uncertainty instead of 256. For a 12-word phrase, four unknown words are 44 bits less 4 of checksum, leaving 40 instead of 128. For scale, the standard written specifically for splitting wallet secrets, discussed next, requires at least 128 bits of entropy so that the one small reduction it permits stays impractical. A split 12-word phrase leaves whoever finds one location 40.
What Shamir shares do differently
SLIP-39 standardises Shamir’s secret sharing for wallet backups. It splits a secret into a number of shares such that a chosen threshold of them reconstructs it, and its abstract states that knowing fewer than the threshold does not leak information about the secret. The mechanism is geometric. The secret is encoded as one value of a polynomial whose degree is below the threshold, and each share is another point on the same curve. Three points define a curve of degree two exactly; two points are consistent with every possible secret. Holding two shares of a three-of-five set tells you nothing that a guess would not.
The spec is candid about the one qualification. Each set carries a 4-byte digest so that a wrong combination of shares can be detected, and an attacker holding one share fewer than the threshold could in principle use it to rule out candidates, cutting 32 bits from the secret, but only by first searching every possible value. That is the reason for the 128-bit floor.
In practice a SLIP-39 share is 20 words for a 128-bit secret or 33 words for a 256-bit one, drawn from the standard’s own 1,024-word list, not BIP-39’s. Every share in a set begins with the same two words, so shares that belong together can be recognised, and each carries a Reed-Solomon checksum guaranteed to detect up to three errors. Shares can also be arranged in two levels, groups with their own thresholds inside an overall one. The spec’s worked example is a holder who can recover alone with two shares she keeps herself, or without them using three of five friends together with two of six family members.
The costs are as concrete as the benefits. The two standards are, in the spec’s own words, for the most part not compatible, a point its design rationale explains at length. An existing BIP-39 phrase can be converted, but only into 59-word shares, and the spec’s advice is to move funds to a new wallet backed up with SLIP-39 from the start, which also retires any forgotten copy of the old phrase. Recovery needs software or a device that implements the standard. And at the moment of recovery the shares are brought together and the whole secret exists on one device: the threshold protects the backup, not that moment.
Multisig splits authority instead of the secret
Multisig reaches a similar goal by a different route. There is no single secret to cut up. A multisig wallet has several independent keys, each with its own phrase, and the chain itself demands that a set number of them sign. Bitcoin’s M-of-N transaction form lists every public key in the script and requires the stated number of valid signatures. Nothing is ever reassembled, so there is no moment when everything exists in one place; each key signs on its own device.
That changes what each backup is. One phrase from a two-of-three wallet is a complete secret for one key and useless for spending on its own, which is what people wanted from cutting a phrase up. The catch is that the spending script names every public key, so rebuilding the wallet needs all of them and the threshold, not merely two private keys. The configuration has to be backed up alongside the phrases. How that works, and what it costs in complexity, is in multisig wallets explained.
A lighter version of the same idea is a BIP-39 passphrase kept apart from the phrase. BIP-39 notes that every passphrase produces a valid wallet and only the right one produces yours, so the phrase alone opens a different, normally empty, wallet. The passphrase then becomes something that can be lost with the same consequences, and it protects only as well as it resists guessing by whoever holds the phrase. What a seed phrase passphrase does covers both sides.
Photos, cloud notes and email
All three make a copy you do not control. A photo in a camera roll that syncs, a note in a cloud account and an email to yourself are each protected by that account’s login, its recovery questions, the provider’s security, and every device signed into it. MetaMask’s own guidance names a cloud-saved document and an email labelled as the recovery phrase as easily hacked places, and tells users to keep the phrase offline. How crypto wallets actually fail starts with this category, and whether a seed phrase can be hacked covers the remote routes in detail.
Typing the phrase into anything belongs on the same list. A phrase typed on a networked computer “to save it somewhere” has passed through that machine’s keyboard, its clipboard and possibly its backups, and has turned a cold wallet’s root back into a hot one.
The password manager question
The honest version is a trade rather than a verdict. A password manager is an encrypted copy held on networked devices and, usually, on a provider’s servers. It survives a house fire and resists casual discovery, and some products add a route for someone else: Bitwarden’s emergency access, for example, lets a trusted contact request a vault that is opened to them once the holder approves or a waiting period expires. Its encryption is strong if the master password is.
Against that, the phrase now exists wherever the vault is unlocked, so malware on any of those devices can read it, and the master password has in effect become the key to the wallet. A provider’s breach does not need to break the encryption to matter. LastPass’s December 2022 incident notice said an attacker had copied a backup of customer vault data, with secure notes among the encrypted fields, and warned that the attacker “may attempt to use brute force to guess your master password”. An encrypted copy in someone else’s hands can be attacked offline, with no lockout and no time limit.
MetaMask’s position is categorical: its support pages say online storage, password managers included, is not secure for a recovery phrase. The decision turns on which you rate as the larger risk: a physical copy being found or destroyed, or every device and account that can ever open the vault.
Checking the arrangement you have
Four questions test any setup, whichever of the options above it uses:
- Can the copy you would actually use be read correctly? The only proof is a restore, done before the wallet holds anything, as your first crypto wallet describes.
- Who can reach each copy, now and in five years? Count people, not locks.
- What happens if one location is lost entirely? If the answer is that the funds go with it, the arrangement is a single copy with extra steps.
- Would anyone else know it exists? A backup nobody can find serves you and nobody after you.