Hot and Cold Is the Wrong Question
Why the specie market’s custody shorthand misprices modern wallets, misplaces capacity, and what should replace it.
BRIEF
For a decade, the specie market has sorted digital asset custody by temperature: cold storage earns deep limits, hot wallets earn small sublimits. But temperature only says whether keys are online, not who can move the funds or how. Losses have moved to the approval process. Bybit’s $1.5 billion left a cold wallet. I propose a better question: what is the smallest set of failures that moves the funds? A High-Risk Wallet is any custody system whose design has a single point of failure, tested through five lenses, with shared dependencies priced rather than excluded. That lets capacity follow controls, not labels.
On February 21, 2025, attackers took about $1.5 billion in Ethereum assets from Bybit. It came out of a cold wallet.1
That detail should bother anyone who prices custody risk. For a decade, specie syndicates have treated “cold” as safe and “hot” as dangerous. Cold storage earns deep limits; hot wallets earn small sublimits, if they’re covered at all. Bybit’s cold wallet was a multisig requiring three signatures,2 and its signers approved transactions on hardware wallets.3 By the London market’s working shorthand, it was about as safe as custody gets. But when what signers see doesn’t match what they sign, $1.5 billion can walk straight out the vault door.
Bybit absorbed the loss itself. It kept withdrawals open and restored its reserves within days using emergency loans and purchases, so client assets stayed fully backed.4 Someone always pays for a loss like this. The question is whether it was priced in advance.
As in traditional banking, securing digital assets comes down to who can move them and how. “Cold” answers neither question. It only tells you whether the keys are online — call it a wallet’s temperature.
Why the shorthand made sense
The distinction was right when it was born. Early custody meant a key on an internet-connected laptop or a key on paper in a safe, so connectivity really was the exposure. Specie underwriters mapped it onto a world they knew: cold storage was the vault, the hot wallet was the cash drawer. Insure the vault deeply and the drawer lightly. The technology moved on. The shorthand didn’t.
What has changed
Three things broke the link between temperature and risk.
The key stopped being a thing. In multi-party computation (MPC) and threshold-signature custody, the private key is never assembled anywhere. Several parties hold shares and sign jointly, so there’s no whole key on any machine to steal.
Control moved into software. Policy engines, transaction screening, and value and velocity limits mean a “hot” wallet can refuse a transfer that breaks a rule, no matter who signs it. Meanwhile a “cold” wallet can be a smart contract whose logic an attacker can replace.
Losses moved to the approval process. Attackers rarely need to steal a key anymore; they need an approval. In February 2025, Bybit’s signers approved a forged instruction on a compromised interface, and the keys never left cold storage.2 In September 2026, according to Bitget’s CEO, attackers compromised a backend system in its wallet infrastructure, fed spoofed transaction data into its authorization process, and moved more than $350 million out of its hot and warm wallets. She said private key compromise had been ruled out; a full forensic report hasn’t yet been published.5 One cold wallet, one hot, and, on early accounts, the same failure: the approval process trusted information it couldn’t verify. Bitget’s cold wallets were untouched; Bybit’s weren’t. The difference wasn’t temperature.
The better question
If temperature isn’t the variable, what is? Every large custody loss I’ve studied reduces to one question: what is the smallest set of failures that moves the funds?
If the answer is one, the wallet has a single point of failure, and whether it’s hot or cold is irrelevant. If it takes several independent failures, the wallet is resilient, even online. Offline keys still matter: they close one path to the funds. But they’re one control, not a category.
The word doing the work is independent. Safety scientists picture defenses as slices of Swiss cheese: each has holes, and a loss happens only when the holes line up.6 But slices cut from the same block have holes in the same places. Five signatures that all run through one office, one laptop or one screen are one slice, not five.
The CryptoCurrency Security Standard (CCSS) already asks versions of this question across key generation, storage and usage.7 What the insurance market has lacked is a coverage definition that asks it too.
Defining a High-Risk Wallet
A note on scope first. This is an argument about taxonomy: the categories the market uses to sort wallets before underwriting begins. It isn’t a substitute for underwriting. Every risk should still get the full process: a CCSS-based review, a single-point-of-failure analysis of the whole custody system, independent penetration tests and security audits, and an experienced underwriter’s judgment. The narrower question is where a wallet starts. Hot and cold sort wallets by whether their keys are online. The definition below sorts them by single points of failure.
I’ve been working on a definition that replaces hot and cold with that question. A High-Risk Wallet is any custody system whose own design has a single point of failure, tested through five lenses. Here’s the shape of the logic. It’s a working draft; the wording will follow the logic, not the other way round.
Read the lenses again and you’ll notice they ask one question in five places: can one failure, on its own, move the funds? Each line describes a way a single failure could reach the funds without anything else going wrong.
One building (a). Keys and backups in the same facility mean one fire or one raid reaches everything.
One weak key (b). A key is only as strong as the randomness used to create it. If that randomness is flawed, an attacker can work the key out without touching any device. And if the complete key ever sits on a single weak device, stealing that device is enough. This is also where MPC earns credit, and only here. A threshold scheme that never assembles the key, with independently controlled parties behind every transaction, clears lens (b) even though it’s online. It clears nothing else.
One account or one person (c, d). A single administrator who can rewrite the limits, or a single employee who can approve alone, defeats every other control at once.
One screen (e). This lens deserves the most attention, because it’s where the largest recent losses sit. As at Bybit, and on early accounts at Bitget, when everyone and everything in the approval chain checks the same compromised source, several approvals add up to one. That’s why the lens covers changes to a wallet’s rules as well as payments, and why the word that matters is independently: the terms must come from a device or channel the attacker can’t reach. On this test, Bybit’s wallet was High-Risk long before February 2025.
What the test prices instead
Some things the test leaves out on purpose. Every technology has common-mode failures: shared parts that can fail everywhere at once. An insured’s signers may all use the same brand of device, the same wallet software or the same cloud provider. Treating every shared dependency as a single point of failure would rule out almost every wallet. So, in my view, the test shouldn’t exclude them. They should be underwritten and priced.
Property insurers solved this problem long ago. They don’t refuse to write buildings on the coast; they track how much they’ve written there, and price and cap the concentration. Technology dependencies work the same way, at two levels. Within a single risk, a shared dependency means fewer truly independent controls than the submission suggests, and the price should reflect that. Across a book, it means one vendor compromise could reach many insureds at once. The malicious code behind the Bybit theft sat inside an interface used by many organizations. It was aimed at one target, but it didn’t have to be.8
In my day job, we track these dependencies across our portfolio and surcharge concentration, the way a property carrier surcharges a crowded coastline. My view is that losses starting with a third-party provider belong outside the base cover, with a buy-back priced on exactly these dependencies. The test draws a bright line around single points of failure; pricing handles the shared roots that run beneath every line.
What a better definition unlocks
A better taxonomy doesn’t replace underwriting, but it changes what underwriting can do. Four things follow.
Capacity. Capacity providers ask two questions of any risk: how bad can one loss get, and how many losses can happen at once? Under hot and cold, a hot wallet starts with a small sublimit, however well it’s controlled. Under a single-point-of-failure test, a wallet that clears the definition starts out eligible for full consideration, and its limit follows the rest of the underwriting.
Severity controls show what that underwriting looks for. Suppose a wallet can move no more than $5 million an hour, enforced outside the wallet’s own rules so an attacker can’t simply switch it off, and the insured can detect and freeze a compromise within four hours. One compromise is then bounded near $20 million, whether the wallet is online or not. A well-controlled “hot” wallet with controls like these can responsibly carry limits once reserved for cold storage. The second question, how many losses can happen at once, is what dependency tracking is for. Together, the two let capacity follow controls rather than labels. Sizing both properly is a topic for its own post.
Incentives. The hot/cold model quietly rewards concentration: if cold means safe, the rational move is to put as much as possible behind one cold control set. Bybit shows what that concentration costs when the control set fails. A control-based test rewards the opposite: demonstrable independence, bounded exposure, and limits that grow as evidence of good controls accumulates.
Certainty. A testable definition tells the insured, the broker and the capacity provider, before a loss, whether a wallet is High-Risk. Bright lines are what make deep capacity comfortable.
Regulation. Some rules already set requirements by temperature: Hong Kong’s SFC mandates cold-storage ratios for licensed platforms and sets compensation requirements by storage type.9 The SFC has since called for technology-neutral, outcome-based custody standards,10 and a definition built on single points of failure is one way to deliver them.
A working draft, not a final answer
This piece argues for one change: sorting wallets by single points of failure rather than by whether their keys are online. Each of the other parts of my working draft, including transfer controls, device standards, and attacks upstream and downstream of the custody system, deserves its own discussion.
Even the part argued here is best treated as a hypothesis. A definition describes how a wallet is designed, losses happen where reality departs from that design, and every new post-mortem is a test of whether the definition would have caught the failure. This one has already failed two of my own tests while I wrote it: lens (e) had to grow to cover changes to a wallet’s rules, and a catch-all exclusion I drafted turned out to belong in pricing. I expect future losses, and readers, to find more opportunities for improvement.
An invitation
Language like this only works if it becomes a market standard. If you underwrite, broker or provide capacity for custody, test it and tell me where it breaks on LinkedIn or X. And if you take one thing from this piece, make it a question to ask of every custody submission: what is the smallest set of failures that moves the funds?
The market has priced temperature for a decade. It’s time to price control.
Sources
- Sygnia: Bybit investigation
- NCC Group: In-depth technical analysis of the Bybit hack
- Blockstream: What the Bybit exploit reveals about enterprise custody
- CNBC: Bybit says it fully replenished reserves after record $1.5 billion hack
- CoinDesk: Bitget's $351 million hack happened via spoofed transfers, not private keys
- James Reason, “Human error: models and management,” BMJ (2000)
- CryptoCurrency Certification Consortium: CCSS v9.0 release
- Wilson Center: The Bybit heist: what happened, what now
- Davis Polk: Hong Kong licensing regime for virtual asset trading platforms comes into force
- SFC: ASPIRe regulatory roadmap for Hong Kong's virtual asset market
I work at a Lloyd’s coverholder. Views are my own, not those of my employer or any carrier. Nothing here is an offer of coverage or a statement of how any policy will respond.