Keys, wallets and custody, without the jargon

A wallet does not contain assets. It contains keys, and it constructs and signs messages with them. Everything that follows from that sentence, including why a lost phrase is unrecoverable and why an exchange balance is a different kind of thing altogether, is the practical core of holding a digital asset.

Keys, phrases and addresses

The arithmetic is one-directional. A private key is a very large random number. From it, a public key is derived, and from that a shorter address that can be shared safely. Deriving the address from the key is trivial; going backwards is not feasible with any known method, and that asymmetry is the entire security model. Nothing about it depends on a company staying in business.

A seed phrase of twelve or twenty-four words encodes the single random secret from which a wallet derives many keys deterministically. That is why the phrase alone restores an entire wallet onto new hardware, and why the phrase is the asset in any meaningful sense. Anyone who reads it controls everything derived from it, immediately and irreversibly, and no issuer, developer or exchange can intervene afterward. Photographing it, typing it into a website, and storing it in an account that syncs to a compromised device are the three routes by which most self-custodied assets are lost.

Addresses themselves carry no identity and no error correction beyond a checksum. Two addresses can differ in one character and belong to unrelated parties, which is why attackers work so hard to get a wrong address in front of someone at the moment they are copying one.

What the software actually does

Wallets differ mainly in where the key sits when it signs. A hot wallet keeps the key on an internet-connected device, which is convenient and exposed to whatever else that device runs. A hardware signer keeps the key in a separate chip and displays the transaction on its own screen, so malware on the computer can propose a payment but cannot sign one, and cannot silently alter a recipient the user has verified on the device itself. Cold storage generalizes this to keys that never touch a connected machine at all.

Beyond single keys there are two arrangements worth knowing. A multisignature wallet requires several independent keys to authorize a transfer, so one compromised key or one lost key is survivable, which is why it is standard for organizational treasuries and protocol funds. An MPC wallet splits a single key mathematically so that no complete key exists in one place, which suits institutional operations but transfers part of the trust to the vendor's implementation. Account abstraction adds programmable rules to the account itself, including spending limits and social recovery, at the cost of depending on more code.

A separate mechanism causes a large share of real losses and is easy to overlook. Interacting with a contract usually requires granting it permission to move a token balance, and those approvals persist after the interaction ends. An approval granted years ago to a contract that is later compromised remains exploitable until it is revoked, which is why periodically reviewing outstanding approvals matters as much as protecting the phrase.

Three custody models, three different loss lists

Under self-custody, the holder has the keys. Nobody can freeze the balance, and nobody can help after a mistake. The realistic threats are key loss, phishing that persuades the holder to sign something they misread, malicious approvals of the kind just described, and address poisoning that plants a lookalike address in transaction history.

Under exchange or broker custody, the provider holds the keys and the customer holds a database entry plus a contractual claim. The threats change entirely: compromise of the provider's hot wallets, commingling of client funds with the firm's own, reuse of customer assets as collateral, account freezes, and outright insolvency. The 2022 failures of Celsius, Three Arrows Capital and FTX are the reference cases, and in each of them customers looking at a balance on a screen turned out to be unsecured creditors of a company. Proof of reserves partially addresses this by demonstrating control of addresses, but a reserve statement without an equally verified liability side does not establish solvency, and that limitation is structural rather than a matter of effort.

Under regulated third-party custody, a qualified custodian holds assets under a specific legal regime, usually with segregation requirements, audits and insurance of defined scope. An exchange-traded product sits on top of this: the investor holds a security, the custodian holds the asset, and the exposures are issuer, custodian, fee and tracking risks rather than key management. It removes one category of loss by introducing counterparties, which is a trade rather than an unambiguous improvement, and the terms of that trade are set out in the product's own documents.

The decisions that actually reduce loss

Most losses are not exotic cryptography failures. They are a phrase stored where someone else could read it, a single device holding both the key and the browser that met a phishing page, a signature approved without reading what it authorized, or an entire balance left with a third party whose books were never verifiable by anyone outside it. Testing a recovery phrase on a spare device before it matters, separating the signing device from the browsing device, revoking stale approvals, and treating any request to enter a phrase into software as hostile address the large majority of realistic scenarios.

The awkward part is rarely stated. Self-custody moves the failure point from an institution to a household, including the question of what happens if the holder dies or becomes incapacitated. Institutions have succession procedures, account recovery and legal process; a phrase in a drawer has none of that unless someone deliberately builds it, and inheritance arrangements for keys are genuinely difficult because any scheme that lets a family member recover the asset also describes a way for someone else to.

The incidents record documents how custody has actually failed in practice, and the risk pages set out the categories in a consistent form across assets. On the data side, addresses holding a balance gives a rough sense of how widely an asset is held directly on-chain, while coins held by exchange-traded products shows how much sits inside the regulated custody model instead.