Read the post-mortems and a pattern emerges quickly. Exchanges are almost never broken by cryptography. They are broken by operations.
An admin account with more permissions than the role needed. A hot wallet holding a month of withdrawals because rebalancing was manual and someone was on leave. A withdrawal flow one person could push through alone.
Key management is the whole game
Every other security decision is downstream of how keys are held.
Single-key custody has no place in an exchange. One compromise, total loss, no recovery.
Multisig requires several keys to authorise a movement. Solid, transparent on-chain, and operationally rigid — signer changes mean on-chain transactions, and support differs across chains.
MPC splits a key into shares so that no party ever holds it whole, and produces an ordinary-looking signature. Cheaper on-chain, chain-agnostic, and much easier to rotate signers. The trade is that the security lives in the protocol implementation rather than in something the chain enforces, which is why the implementation has to be reviewed properly.
For most venues MPC is the better fit. Either is enormously better than what the incident reports keep describing.
Hot wallet exposure should be enforced, not documented
Every exchange needs hot wallets. The question is how much sits in them and who notices when that changes.
The answer that works is a rebalancing daemon: continuously sweeping excess into cold storage, holding only what near-term withdrawals require, and alerting when exposure crosses a threshold.
A policy saying hot wallets should hold under a certain amount is a document. A daemon enforcing it is a control. Incidents keep finding the difference.
Withdrawals are the attack surface that matters
An attacker who compromises your systems wants one thing. Everything about the withdrawal path should reflect that.
- Multiple humans on anything above a threshold, from different systems.
- Address allowlisting with a delay before a newly added address can receive.
- Velocity limits at account and system level, halting automatically on anomalies.
- Circuit breakers any on-call engineer can pull immediately.
The delay on new addresses is the single highest-value control here and the one users complain about most. It has saved more balances than any other measure on this list.
Admin access is where insiders and attackers meet
Most exchange admin consoles start permissive and stay that way, because scoping permissions is tedious and nothing breaks if you skip it.
What good looks like: every action attributable to a named individual, permissions scoped to the minimum each role needs, no shared accounts, hardware keys for anything sensitive, and a full log of who did what.
The log matters most. When something goes wrong, the question is always what happened, and a system that cannot answer leaves you reconstructing from fragments.
Audits, and who does them
We coordinate independent third-party firms and the report comes from them. We do not audit our own code — that is a conflict of interest dressed up as a service.
Two independent firms is the baseline for anything holding meaningful value, plus formal verification on components where a bug is unrecoverable. Budget for remediation, not just for the audit: finding issues is the easy half.
Alongside contract audits, an infrastructure penetration test against a live-like environment catches a different class of problem entirely.
Rehearse the incident
Every exchange has an incident response plan. Most have never run it.
The rehearsal is where you discover that the on-call engineer cannot actually reach the circuit breaker, that nobody knows who authorises a halt, and that the communications draft assumes facts you will not have for two hours.
Run it as a drill before launch, and again periodically. The plan is not the artefact. The rehearsal is.
Common questions
How do crypto exchanges get hacked?
Overwhelmingly through operations rather than cryptography. Compromised admin credentials, hot wallets holding far more than they should, withdrawal approval flows a single person can push through, and insider access that was never scoped down. The cryptography in a modern exchange is rarely the weak point. The people, the processes and the permissions around it almost always are.
How much should an exchange hold in hot wallets?
Only what withdrawals actually need over a short horizon, with automated rebalancing pulling the rest into cold storage continuously. The number is specific to your withdrawal patterns rather than a fixed percentage. What matters is that the limit is enforced by software rather than policy, because a policy is a document and a daemon is a control.
What is MPC custody and is it better than multisig?
MPC splits a private key into shares held by separate parties, so a signature is produced without any single party ever holding the whole key. It differs from multisig in that the chain sees one ordinary signature, which makes it cheaper and chain-agnostic. Both are substantial improvements on single-key custody. MPC tends to win on operational flexibility and cross-chain support; multisig wins on on-chain transparency.
What should an exchange do before launch to be secure?
Coordinate independent third-party audits of the contracts and the infrastructure, run a penetration test against the live-like environment, scope admin permissions down to the minimum each role needs, put withdrawal approvals behind multiple humans, and rehearse the incident response rather than writing it down and filing it. The rehearsal is the part most teams skip and the part that decides how a real incident goes.
Does Corum8 handle exchange security?
We build the infrastructure securely and coordinate independent third-party audit firms rather than auditing our own work. That covers key management and MPC or HSM custody design, withdrawal approval flows, admin permission scoping, monitoring and anomaly alerting, and the operational runbooks your team uses when something goes wrong.
