Every time a major exchange collapses, the same conversation restarts. Centralized exchanges are dangerous, people say. Go self-custodial. Move on-chain. The reality is more precise than that, and the imprecision costs traders real money.
Here is the honest version.
The Framing Everyone Gets Wrong
The standard CEX vs DEX debate runs on one axis: custodial versus non-custodial. That framing misses the actual question, which is not who holds the keys but what failure mode you are exposed to, and whether you can survive it.
A centralized exchange trade is a database update. Your balance is an entry in the exchange's internal ledger. The actual coins sit in a commingled pool that the exchange controls, can lend out, can lose, and in several documented cases has rehypothecated without user knowledge. A withdrawal is the only point at which on-chain settlement occurs. Until then, you hold an IOU.
A decentralized exchange trade is a direct call to a smart contract using your own keys. The protocol cannot freeze you, cannot reverse your transaction, and cannot help you if something goes wrong. Both of those properties are simultaneously the reason DEXs exist and the reason self-custody alone is not a safety strategy.
The real choice is between human failure modes and code failure modes. Both are catastrophic. They just fail differently.
How CEXs Actually Fail
Rehypothecation is the default, not the exception.
Most centralized exchange terms of service include language granting the exchange discretion to lend, pledge, or otherwise use deposited assets. One of the largest lending platforms to collapse in the 2022 cycle had terms explicitly stating the platform held ownership rights over deposited assets and could rehypothecate them without restriction. When it filed for bankruptcy, a judge ruled that depositors were unsecured creditors, not beneficial owners. That distinction cost hundreds of thousands of users a substantial portion of their savings.
Proof of reserves proves assets, not solvency.
Merkle-tree proof-of-reserves attestations confirm the exchange controls certain wallet addresses at a snapshot in time. They do not confirm that those assets are not borrowed from a sister entity specifically for the snapshot period. They do not confirm off-chain liabilities, outstanding loans, or cross-entity exposures. Only a full proof-of-solvency combining assets, liabilities, and independently verified off-chain obligations is meaningful. Almost no exchange provides this.
The WazirX case from July 2024 is the clearest example. The exchange had published a proof-of-reserves attestation claiming approximately 500 million dollars in assets. One month later, a 235 million dollar hack revealed that nearly half of those assets were accessible to attackers through a compromised multisig interface. Proof-of-reserves did not warn any user because it was not designed to.
The largest single theft in crypto history hit a centralized exchange.
In February 2025, attackers stole approximately 1.46 billion dollars from a major exchange by compromising the user interface of the multisig wallet management tool the exchange was using for cold storage operations. The exchange's signers held hardware wallets and followed their standard signing process. They signed a malicious transaction anyway because the interface displayed legitimate-looking data while submitting a different payload. This is called blind signing, and it remains unresolved across the industry.
The exchange covered the full loss through its existing reserves plus emergency bridge loans from institutional counterparties and recovered its proof-of-reserves ratio within 72 hours. Users were made whole. Most exchanges could not absorb a loss of that size.
Withdrawal freezes are a policy tool, not an edge case.
Exchanges operating under regulatory oversight can and do freeze accounts without user consent under AML, KYC, or law enforcement requests. In documented cases this has lasted months. Accounts flagged for receiving funds from unrelated parties have been frozen pending investigations that can take longer than the regulatory timeline suggests. One major exchange paid 505 million dollars in penalties in early 2025 for AML failures without preventing the underlying pattern. For any user whose jurisdiction is subject to active regulatory enforcement, the ability to exit a position on a centralized exchange is conditional, not guaranteed.
Order books are partly synthetic.
Centralized exchanges with proprietary trading desks have structural visibility into order flow before execution. Spoofing and wash trading remain measurable on mid-tier and offshore venues. A futures contract that shows a 138% price spike in 30 minutes on one exchange while trading normally on all others is not a market event. It is an operational event. Documented cases of this resulted in trade rollbacks and account freezes in 2025, with the exchange deploying its protection fund to cover affected traders. The intervention resolved the immediate issue. It also demonstrated that the exchange has unilateral power over completed trades.
How DEXs Actually Fail
The smart contract is only one attack surface.
Most high-profile DEX losses in 2024 and 2025 did not come from bugs in the core swap logic. They came from oracle manipulation, frontend compromise, bridge exploits, and governance attacks. Understanding that the primary contract is audited does not mean the system is secure.
Oracle manipulation was the second most damaging on-chain attack vector in 2024, accounting for over 52 million dollars across 37 separate incidents that year. The mechanism is consistent across cases: an attacker takes a flash loan, manipulates the spot price of a thinly liquid token on an AMM, triggers a lending protocol or vault that reads that price as its oracle, and extracts collateral at the artificially inflated valuation. The contracts execute exactly as designed. The design assumption about price feed integrity was wrong.
Cross-chain bridges are the dominant loss category.
Bridges have accounted for over 2.8 billion dollars in stolen funds since 2022, representing roughly 40% of all on-chain hacked value. The fundamental problem is that bridging requires a trust assumption at the relay layer, whether that is a multisig threshold, a validator set, or a proof system. Each of those components has a documented compromise vector. In 2025, cross-chain bridges were responsible for over 50% of all crypto theft by dollar value in the first half of the year. Users who moved funds between chains through lesser-audited bridges and assumed the on-chain settlement was equivalent to mainnet security were wrong.
Frontend attacks break the self-custody guarantee.
In May 2025, attackers compromised the domain registrar of a major DeFi protocol, redirected the curve.fi DNS record to a wallet drainer, and stole approximately 3.5 million dollars from users who visited the legitimate URL. The underlying smart contracts were never touched. This was not a DeFi failure in the technical sense. It was a Web2 failure that defeated self-custody entirely because users were interacting with a spoofed interface.
This is the most underappreciated structural gap in decentralized finance. The protocol is on-chain. The frontend is not. Any DEX accessible through a standard web browser is partially dependent on DNS registrars, CDN providers, and JavaScript integrity. A user who takes no additional precautions is trusting all of those layers every time they connect their wallet.
The "non-custodial" label covers a range of actual risks.
In March 2025, an attacker manipulated a low-liquidity perpetual market on a leading on-chain derivatives platform by opening a large position, then self-liquidating it into the platform's community liquidity vault. The vault's unrealized losses reached approximately 13.5 million dollars. The platform's validator set, consisting of a small number of validators, voted within two minutes to delist the token and force-settle all positions at a specific price, converting the vault's loss into a modest gain. Users with opposing positions received settlement at the validator-determined price, not the market price.
This sequence was operationally sound from the platform's perspective. It also demonstrated that validator intervention on a nominally decentralized venue can produce outcomes indistinguishable from a centralized exchange administrator halting a market. The difference between a centralized exchange freezing a trade and a small validator set force-settling a position is procedural, not categorical. Both venues have intervention capacity. The question is what triggers it and who holds it.
Blind signing on hardware wallets does not mean you are protected.
Several of the largest multisig exploits in 2024 and 2025 involved hardware wallets. The WazirX attack, the Bybit attack, and the Radiant Capital exploit all occurred while hardware wallets were in use. In each case, the device signed a transaction it could not decode in full because the payload exceeded or bypassed the device's clear-signing capability. The device confirmed that a transaction was being signed. It could not confirm what the transaction was doing at the contract level. Until every wallet interaction in DeFi produces a fully human-readable description of every state change on the device screen, hardware wallet ownership does not guarantee protection from UI-layer attacks.
What the Data Says About Migration
DEX market share has been rising structurally. The ratio of DEX to CEX spot volume reached a new high of 21.2% in November 2025 according to CoinGecko, with a peak of 27.9% recorded in June 2025. On-chain perpetual futures volume crossed one trillion dollars per month for two consecutive months in late 2025. The total value locked in DeFi reached 237 billion dollars in the third quarter of 2025, a record. Unique active DeFi wallets hit 27.3 million in May 2025.
Hardware wallet sales reflect the same shift. One major hardware wallet manufacturer reported revenues in the triple-digit millions in 2025, up from 70.9 million dollars in 2024 and 36.7 million dollars in 2023. Cold wallet ownership among retail users rose 34% year over year entering 2025. Non-custodial wallets now represent approximately 59% of all crypto wallet usage globally.
These numbers tell a real story about structural preference change. They do not tell a simple story about safety. The migration to on-chain activity has been accompanied by 3 billion dollars in crypto theft in the first half of 2025 alone, already 50% above the full-year 2024 total. Daily active DeFi wallets fell 22.4% in the third quarter of 2025 despite record total value locked, suggesting the growth in on-chain capital is concentrated in fewer, larger participants rather than distributed retail adoption.
The market is moving on-chain. It is not uniformly safer on-chain.
The Honest Risk Map
Neither venue is categorically safer. Each one is safer than the other for specific use cases, and understanding that distinction is more valuable than any tribal take.
Long-term spot holdings belong in self-custody with a hardware wallet and clear-signing firmware. Exchange insolvency risk compounds over time. A position held on a centralized exchange for three years has been exposed to that exchange's operational, regulatory, and counterparty risk for the entire duration. A position in a self-custodied cold wallet has been exposed to seed-phrase management risk and hardware failure risk. For most users holding liquid assets with no active trading need, the cold wallet risk profile is lower. The billion-dollar exchange compromise in February 2025 did not affect any user's personal wallet, only the exchange's own custody infrastructure.
Active leveraged trading belongs on venues with deep liquidity, functioning insurance funds, and regulatory accountability. On-chain perpetual platforms have made significant progress but remain smaller, more concentrated in their validator sets, and less tested under extreme market conditions than the top centralized exchanges. The on-chain CLOB platforms that lead this category have demonstrated both impressive resilience and concentrated intervention capacity within twelve months of each other.
New token speculation belongs on-chain. Tokens generate on-chain liquidity before they list on centralized venues. Waiting for a centralized listing means trading into a known exit event with reduced information advantage. On-chain access to early-stage liquidity is one area where decentralized venues have an irreplaceable structural advantage.
The Four Security Decisions That Matter Most
First, use a private RPC endpoint. Broadcasting a swap to the public mempool exposes the transaction to sandwich bots before it lands in a block. Enabling a private RPC takes under two minutes and eliminates the primary MEV attack surface for standard swaps. It does not eliminate all on-chain risks but removes the most mechanically reliable one.
Second, revoke token approvals regularly. Every DeFi interaction that requires approval grants the protocol permission to move a specified token amount from your wallet. Infinite approvals granted months or years ago to protocols that have since been compromised or updated represent standing risk. Revocation takes one transaction per approval and costs a small amount of gas. Doing this quarterly removes a category of risk that has been responsible for millions in losses following frontend and protocol compromises.
Third, verify the contract address, not the website. DNS attacks succeed because users trust the URL. The actual security property of a DEX interaction is the contract address being called, not the domain name resolving to it. Before interacting with any significant value, check the contract address against the protocol's official GitHub or audit report. If they do not match, stop.
Fourth, size CEX balances as active float, not savings. A centralized exchange balance should represent the capital you intend to actively trade in the near term. The operational requirement to access fiat ramps and leveraged derivatives is real and CEXs handle it better than any current on-chain alternative. That operational requirement does not extend to long-term storage. The historical loss rate on mid-tier and offshore centralized exchanges is high enough that treating a CEX balance as equivalent to a bank deposit is a category error.
Where This Lands
The CEX versus DEX question has never had a clean answer, and the 2025 data does not provide one. It provides a more granular picture: concentrated risk on centralized venues is real and has cost users billions across multiple well-documented events; distributed risk on decentralized venues is also real and has cost users billions across a different set of well-documented events.
What has changed is that on-chain infrastructure has matured enough to handle most retail use cases, hardware wallet adoption has reached meaningful scale, and the regulatory environment in major jurisdictions has clarified enough that self-custody is a defensible strategy rather than an ideological choice.
The traders who navigated 2025 without major losses were not the ones who picked the right side of the debate. They were the ones who understood that every signature is a security decision, that every custody arrangement is a risk transfer to a specific counterparty, and that the question is never CEX or DEX but always which risk are you willing to own and have you understood it fully before signing.
