Smart-Account Wallets in 2026: What the Radiant Heist Teaches

In October 2024, three engineers at a lending protocol called Radiant Capital approved what their screens told them was a routine transaction. The signatures were valid. The multisig did exactly what it was designed to do. And roughly $50 million walked out the door. The smart contract was never broken; the people signing for it were. That gap, between what an account is told to do and what its owners believe they are approving, is the whole story of smart-account wallets in 2026.

Smart-account wallets are the biggest change in how ordinary people hold crypto since the seed phrase. They turn a passive keypair into programmable code that can batch actions, pay its own gas in stablecoins, log in with a passkey, and recover from a lost device without a scrap of paper. Tens of millions of accounts already run this way, most of their owners never noticing. But moving authorization into code also opens a new place to be robbed, and the largest thefts of the past two years have all happened there. This is a plain-English guide to what a smart-account wallet is, the standards that make it work, what it lets you do, and what the Radiant heist and its cousins teach about where the model fails. Ether changed hands near $2,490 as this published, drifting with the rest of the market ahead of a US inflation print (see our read on the setup into September’s decision week), but the wallet story runs on a longer clock than any single candle.

What a smart-account wallet actually is

Ethereum has always had two kinds of account. An externally owned account, or EOA, is controlled by a single private key, and the protocol fixes its rules: one key signs, one signature is enough, and the account can do nothing on its own. A contract account is controlled by code. For a decade, every ordinary wallet (a fresh MetaMask, a Ledger, a Coinbase Wallet) was an EOA, which meant your security was only ever as good as one secret phrase written on paper.

A smart-account wallet flips that arrangement. The account itself is a smart contract, so its owner writes the rules that govern it: who may sign, how many signatures are required, what a daily spending cap is, which contracts a temporary key may touch, and how to recover access if a device is lost. The umbrella term for making an Ethereum account programmable is account abstraction, and in 2026 there are three ways to get there. The first, ERC-4337, builds a smart account as a separate contract. The second, EIP-7702, upgrades the EOA you already own in place. The third, native account abstraction, is a set of proposals to bake the whole idea into the protocol so no add-on layer is needed. The table below sketches how they differ.

The two roads: ERC-4337 and EIP-7702

ERC-4337 shipped as a final standard in March 2023 and did something clever: it delivered account abstraction without changing Ethereum’s consensus rules at all. Instead of a normal transaction, a 4337 user signs a UserOperation and drops it into a separate mempool. Operators called bundlers gather those UserOperations, wrap them into a real transaction, and route them through a single shared contract, the EntryPoint, which verifies each account’s rules before executing. A paymaster contract can step in to cover gas. The EntryPoint has moved from v0.6 through v0.7 to v0.8, the last of which added native support for the newer EIP-7702 path. Adoption is not a rounding error: the tracking site BundleBear counts more than 1.27 billion UserOperations across over 66 million accounts, with roughly $14 million in gas sponsored by paymasters, and the standard’s code lives in the public eth-infinitism releases.

The catch with 4337 is that it asks you to create a brand-new account. EIP-7702, activated in Ethereum’s Pectra upgrade in May 2025, closes that gap. It introduces a new transaction type (0x04) that lets an existing EOA delegate to a contract’s code while keeping the same address and the same key. In practice your everyday wallet gains smart-account powers without you moving a single token or memorising a new address, and the delegation is revocable: point it back at the null address and you are a plain EOA again. That convenience is why uptake has been fast. BundleBear records more than 240 million cumulative authorizations and over 56 million live delegations, though, as we will see, a large share of the raw authorization count comes from automated sweeper contracts rather than real users. The short version: 4337 is a parallel system you opt into, while 7702 is an upgrade to the wallet you already hold.

What smart accounts let you do

The point of all this plumbing is a set of features an EOA simply cannot offer. The first is gasless transactions: through a paymaster, an app can cover your network fee or let you pay it in USDC, so a newcomer never has to acquire ether just to make a first move. Circle’s paymaster, for example, lets an ERC-4337 account pay gas in USDC on chains like Base and Arbitrum.

The second is batching. A smart account can approve a token and swap it in a single atomic transaction, so either both steps happen or neither does; that alone removes a class of stuck approvals and half-finished trades. The third, and the most consequential for gaming and AI, is the session key: a scoped, temporary signer that can act within limits you set (spend up to this much, only on this contract, only until Friday) without a popup for every move. That is what lets an on-chain game feel like a game and an autonomous agent transact on your behalf under a leash. Standards such as ERC-7715 and MetaMask’s Advanced Permissions formalise how a wallet grants those bounded permissions. Add spending limits, allow-lists, automated recurring payments, and one-click multi-step actions, and the smart account starts to behave less like a keychain and more like a bank account whose rules you wrote yourself.

Those bounded permissions are why smart accounts matter for gaming and for the fast-growing class of on-chain AI agents. A blockchain game that made you sign a wallet popup for every sword swing would be unplayable; a session key lets the game act inside a budget you approved once, then expires on its own. The same pattern lets an autonomous agent pay for data or compute on your behalf without ever holding your keys outright. That is a genuine shift in who, or what, is allowed to move money out of your account, and it is why session keys are among the most requested smart-account features from studios and agent builders alike.

Passkeys, session keys, and social recovery

The feature most likely to change everyday behaviour is the one that kills the seed phrase. Because a smart account defines its own signing rules, it can accept a passkey, the same WebAuthn credential that already unlocks your phone with a face or a fingerprint, generated and held inside a hardware secure enclave. Passkeys use the P-256 (secp256r1) curve rather than Ethereum’s native secp256k1, and verifying that curve on-chain used to cost roughly 300,000 gas. A precompile called RIP-7212 cuts that to about 3,450 gas, an almost hundredfold saving, which is what makes passkey signing economical on the layer-2 networks where most smart accounts live. The FIDO Alliance reports billions of passkeys already in use, so the on-ramp is enormous.

The other half of the story is recovery. On an EOA, losing the seed phrase means losing everything, forever. A smart account can instead nominate guardians (other devices, trusted people, or an institution) who can jointly help you rotate to a new signer if a key is lost, without ever holding your funds. Argent, now rebranded as Ready, pioneered this social-recovery model years ago, and it is now a common feature. Together, passkeys and social recovery are what let a self-custody wallet feel as forgiving as a login you already trust, which is the single biggest reason ordinary users are ending up in smart accounts at all.

There is a catch worth naming. A passkey is only as portable as the platform that stores it, and today an Apple passkey does not natively move to an Android device; the FIDO Alliance’s credential-exchange formats that would fix that are still being finalised. That makes the sync fabric behind your passkey (iCloud Keychain, a Google account, or a third-party password manager) a new dependency to reckon with, because losing access to it can mean losing your signer. The better smart-account wallets pair a passkey with independent recovery rather than betting everything on one vendor’s cloud.

The wallet landscape in 2026

The market has sorted into a handful of camps. Safe remains the giant of the institutional and treasury world; its Q2 2026 report counted close to 130 million transactions in the quarter and more than 63 million accounts, holding tens of billions of dollars. MetaMask now offers smart-account features to its enormous user base along the EIP-7702 path. Coinbase pushed hardest on the consumer side, turning its Smart Wallet into Base Account with passkey sign-in, and launching Embedded Wallets that keep keys inside a trusted execution environment the company says it cannot access. Behind these front-ends sit the infrastructure providers: ZeroDev’s Kernel, Biconomy’s Nexus, and others whose modular accounts power dozens of apps. The table gives a rough map.

Under the surface, many of these accounts are built the same way: as modular contracts whose features are installable plug-ins. A standard called ERC-7579 defines common module types (validators that decide what counts as a valid signature, executors that carry out actions, and hooks that enforce policy) so that a passkey validator or a spending-limit hook written for one wallet can, in principle, run in another. That modularity is what lets the ecosystem ship new capabilities without every team rebuilding an account from scratch, and it is a large part of why smart-account features have spread across wallets so quickly.

The Radiant Capital heist: an account that did as it was told

Now to the cautionary tale, because it is the clearest lesson in the whole field. Radiant Capital was a cross-chain lending protocol, and its funds were governed by a Safe multisig requiring three signatures out of eleven authorised signers. That is a strong setup on paper. It did not matter.

According to the protocol’s own post-mortem and analysis by the security firm Halborn, the intrusion began in mid-September 2024, when a Radiant developer received a Telegram message from someone posing as a trusted former contractor, asking for feedback on a smart-contract auditing report. The attached ZIP hid a decoy PDF alongside a persistent macOS backdoor that researchers later named INLETDRIFT. This was not a coding error in a contract; it was a targeted social-engineering campaign of the kind we cover in our look at crypto’s supply-chain and phishing attacks. Roughly five weeks later, on 16 October 2024, the trap sprang. Malware displayed correct, legitimate transaction data inside the Safe front-end that the signers were looking at, while entirely different, malicious payloads were sent to their hardware wallets to be signed. The attackers even mimicked routine transaction failures to nudge signer after signer into re-approving, harvesting the three signatures they needed. Then they called transferOwnership and swapped the lending-pool implementation for their own contract.

About $50 million (some estimates put it at $53 million) drained from the Arbitrum and BNB Chain deployments. Cybersecurity firm Mandiant later attributed the attack to UNC4736, a group aligned with North Korea’s Reconnaissance General Bureau. The protocol never recovered. After eighteen months of failed recovery efforts, Radiant’s DAO voted to begin an orderly wind-down on 1 June 2026, with total value locked collapsed to around $1.4 million from a peak near $387 million; users can withdraw through the end of the year. The account had performed flawlessly. That was the problem.

What made the operation so effective was patience. The attackers sat inside Radiant’s environment for roughly five weeks before acting, long enough to learn the team’s signing habits and to time the strike for a routine, expected transaction. The lure itself, a friendly request to review an auditing report, is a hallmark of North Korea-linked crews, who have repeatedly used fake job offers and contractor personas to plant malware on developer machines. None of it required breaking cryptography or finding a bug in the contract code. It required convincing busy people that an ordinary day was still ordinary while the account changed hands beneath them.

Blind signing: the flaw the code cannot fix

The reason a three-of-eleven multisig protected by hardware wallets could be defeated comes down to a single weakness: blind signing. A hardware wallet is meant to be the trusted screen, the device that shows you what you are approving even if your computer is infected. But most hardware wallets cannot decode a complex Safe transaction into human-readable terms. They show a hash, a long string of hexadecimal that means nothing to a person. Faced with that blob, signers approve it because the laptop in front of them, the one running malware, insists everything is normal. The device dutifully signs a payload it cannot explain, and the attacker’s swap goes through.

Security researchers have been blunt about this. Reflecting on the same class of attack, Odysseus, founder of the monitoring firm Phylax, put it plainly: “a hardware wallet is pointless if the transaction is signed on an internet-connected laptop or phone.” Ido Ben Natan, founder of the security firm Blockaid, called the combination of blind signing and custom malware “perhaps the fastest-growing threat in crypto.” The uncomfortable truth is that smart accounts move the point of failure away from the contract and onto the signing screen, and the strongest key management in the world offers no defence if you approve the wrong thing. As our feature on private-key compromise argues, in these operations the theft itself is often the easy part; the days of quiet preparation before it are the real attack. Safe, for its part, rejected the suggestion that its interface was compromised in the Radiant case, framing the loss as a blind-signing error on the signers’ side; that distinction matters, because it means no smart-contract patch can close the hole. Only a change in how humans verify what they approve can.

Radiant in context: WazirX and Bybit

Radiant was not a freak event. It sits in a clear pattern of the largest thefts of the era, and every one of them targeted a smart account without ever breaking its code. In July 2024, the Indian exchange WazirX lost about $230 million when attackers exploited a discrepancy between what a custody interface displayed and what was actually signed, swapping a Gnosis Safe to a malicious contract. In February 2025, Bybit suffered the biggest crypto heist ever, roughly $1.5 billion, when attackers compromised a Safe front-end developer’s machine and injected malicious code into the signing interface; signers saw a normal transfer but authorised a delegatecall that handed the wallet to the attacker. The table lines up the three.

The shared thread matters more than the differences. None of these was a bug in Safe’s contracts; each was a compromise of the software and screens around the account. That is a shift in the economics of crypto theft: as our analysis of cross-chain security costs notes, the money now leaks at the human and interface layer, not through clever contract exploits, and smart accounts, precisely because they concentrate large sums under programmable rules, are the richest target. Safe itself has reframed its mission in response. Lukas Schor, co-founder of the Safe Ecosystem Foundation, described the project as evolving from infrastructure that safeguards value into “a network that can also help verify how value moves.”

The sweeper economy: 7702’s dark side

Retail users face a smaller but far more numerous version of the same problem. Within days of EIP-7702 going live, the trading firm Wintermute observed that more than 97 percent of early delegations pointed at the same copied bytecode: automated sweeper contracts nicknamed CrimeEnjoyor, designed to instantly drain any funds sent to an already-compromised address. That is why the raw authorization counts look so large; they are inflated by bots, not real adoption. Wintermute noted the scheme was barely profitable, moving only a couple of ether across tens of thousands of addresses.

The academic picture is sobering too. An empirical study presented at USENIX Security 2026 analysed millions of EIP-7702 authorizations across seven chains and found a majority of them tied to malicious contracts, with hundreds of distinct malicious contract accounts confirmed by hand and several million dollars in realised and exposed losses. The nuance worth keeping in mind: that majority is measured in transactions, not in unique victims, because attacker contracts are reused so heavily. And none of it is a flaw in the standard. On the brighter side, wallet-drainer losses overall actually fell hard in 2025; Scam Sniffer put the annual drop at 83 percent, to around $84 million, even as individual EIP-7702 phishing cases (one victim lost about $1.54 million to a single malicious batch signature) showed the new attack surface is real.

The more expensive retail losses often have nothing to do with delegations at all; they come from signed token approvals. When you grant a contract permission to move your tokens (an ERC-20 approval, or a gasless Permit or Permit2 signature), a malicious app can drain everything that permission covers with no further click from you. Scam Sniffer’s data has repeatedly put approval-based phishing behind a large share of the biggest individual thefts, which is why revoking stale approvals and reading exactly what a signature authorises matter just as much on a smart account as on a plain EOA.

Clear signing and the ERC-7730 fix

If blind signing is the disease, clear signing is the cure the industry has converged on. The idea is what security people call WYSIWYS, what you see is what you sign: the device you approve on should render a transaction in plain language (send 100 USDC to this address, grant this contract permission to spend this token) rather than a hash. Making that work at scale needs a shared way to describe what a contract’s functions mean, which is the job of ERC-7730, a clear-signing metadata standard now stewarded through the Ethereum Foundation. When a wallet has that metadata, it can show a signer the actual effect of a call, and a Radiant-style swap would look wrong instead of looking like nothing at all.

The limitation is coverage. Clear signing only protects you where the metadata exists and where your device supports it; anything unrecognised still falls back to a hash. That is why, for now, clear signing is a discipline as much as a feature: prefer wallets and hardware that parse transactions, and refuse to approve anything a screen cannot explain. It is slower. It is also the difference between the three Radiant signers seeing an ordinary error and seeing an ownership transfer.

Who pays for gasless

Gasless is not free; it is a billing decision moved out of your view. Paymasters come in three broad shapes. In sponsorship, an app eats the gas as a customer-acquisition cost, the way a game might cover your first hundred moves. In the ERC-20 model, you pay the fee in a token such as USDC and the paymaster takes a markup (Circle’s stablecoin paymaster, for instance, charges around a 10 percent surcharge) to cover conversion and price drift. In the verifying model, an off-chain service decides per request whether to sponsor you. An infrastructure survey by Eco pegs the real cost of a sponsored operation at a few dollars on Ethereum mainnet but a fraction of a cent on layer-2 chains like Base, which is exactly why almost all consumer smart-account activity has migrated to layer 2.

There is a centralisation worry hiding in the convenience. Because 4337 UserOperations flow through bundlers and paymasters, a handful of providers now handle a large share of all smart-account traffic. Vitalik Buterin has warned that leaning on relays and bundlers creates a source of fragility, since a censored or offline relay can leave a user with no path to get a transaction included. The proposed answers, from inclusion-list mechanisms to native account abstraction, all aim to reduce how much you must trust an intermediary just to move your own money.

Exchanges, custody, and the 23-byte designator

Smart accounts also reshape how exchanges and custodians work. When an EOA delegates via EIP-7702, its on-chain code is set to a short 23-byte marker (beginning 0xef0100) that points at the contract it now runs. That single fact creates a screening duty: a deposit address that has been delegated to a sweeper will instantly forward any incoming funds to a thief, so exchanges and custody desks have had to learn to read the designator and flag or block delegated deposit addresses before crediting them. Institutional custodians such as Fireblocks have published security-first guidance urging that accounts delegate only to fully audited contracts, and that high-value holdings stay in setups where a single malicious delegation cannot be catastrophic.

For the ordinary customer, the practical takeaway is that a smart account changes what “withdraw to my wallet” means. Your self-custody address may now carry rules an exchange has never seen, and a screening mismatch can strand a deposit. It is worth checking that any wallet you withdraw to is one whose delegation you understand, which brings the whole discussion back to knowing what your account is actually programmed to do.

Native account abstraction: the road after 7702

The three-way patchwork of EOAs, 4337 accounts, and 7702 delegations is not meant to be permanent. The endgame is native account abstraction, where smart-account behaviour is a first-class part of the protocol and needs no side mempool or external bundler. Two proposals lead. Coinbase and its partners back EIP-8130, a configuration-based design with a new transaction type and an on-chain keystore holding a fixed menu of authenticators (secp256k1, P-256, WebAuthn); Base has been testing it and reports a native USDC transfer costing roughly 63 percent less gas, though at time of writing it remains on a test network with a mainnet date still to be announced. The Ethereum Foundation camp favours EIP-8141, a more flexible “frame transactions” approach that has been scheduled for inclusion in a future upgrade Buterin refers to as Hegota, planned for late 2026 but not as the headline feature. A third effort, Paradigm’s Tempo, takes a deliberately minimal route, with native passkey support but no arbitrary validation logic, on the theory that a smaller surface is a safer one.

Buterin has framed the flexible design as “an omnibus that wraps up and solves every remaining problem that account abstraction was intended to address,” and has said the pieces could come together within a year. Others counsel patience. Ethereum core developer Marius van der Wijden, commenting on the earlier 7702 work, cautioned that “it’s still a very early proposal, so we need to evaluate all the rough edges.” Whichever wins, the direction is clear: fewer intermediaries between you and inclusion, and smart-account features that no longer depend on an add-on layer.

What US rules say, and how not to get drained

For US readers, the regulatory question is mostly settled in a reassuring direction. In April 2026, staff at the Securities and Exchange Commission issued a statement that neutral software that merely lets a user transact from a self-hosted wallet is not, by itself, a broker, provided it does not solicit, route orders, or take custody. In plain terms, the SEC is not treating the smart-account wallet on your phone as a regulated intermediary. Tax treatment does not change either: the Internal Revenue Service taxes gains and losses on disposal the same whether your account is a legacy EOA or a smart contract, a point worth remembering if you hold crypto inside a retirement structure, as our guide to crypto in your 401(k) and IRA lays out.

The line the rules do draw is between your own account and a business acting on your behalf. A self-custody smart account sits outside the broker and money-transmitter regimes, but the moment an exchange or a custodial service holds your keys or routes your orders, know-your-customer, anti-money-laundering, and travel-rule obligations attach to that firm rather than to your wallet. Broader market-structure legislation moving through Congress in 2026 would sharpen those boundaries further, yet the practical point for a self-custody user is unchanged: the code you run is yours to answer for, and so is every signature it makes.

What the rules do not do is protect you from your own signing screen. That protection is a set of habits.

1. Read every transaction on a device that shows the actual action, not a hash. If your hardware wallet only displays a blob, treat clear signing as a hard requirement before you approve.
2. Never blind-sign. If a screen cannot tell you what you are authorising, do not authorise it, no matter how routine the app claims the operation is.
3. Scope session keys tightly: small spending caps, short expiry, and a single contract, so a leaked session key has a small blast radius.
4. Check your delegation. For an EIP-7702 account, look up the 0xef0100 designator on a block explorer, confirm it points at a contract you trust, and reset it to the null address to revoke.
5. Replace the single seed phrase with social recovery or a multisig, and keep guardians or co-signers physically and organisationally apart.
6. For any treasury, hold funds on an audited multisig with a real signing routine: verify raw call data on independent devices, and rehearse the review so a faked transaction error cannot rush a signer.

Frequently Asked Questions

What is a smart-account wallet in simple terms?

It is a crypto wallet that is itself a smart contract, so its owner can program the rules instead of relying on a single private key. That lets it batch actions, pay gas in stablecoins, log in with a passkey, set spending limits, and recover from a lost device, none of which an ordinary key-only account can do.

Is EIP-7702 the same as ERC-4337?

No. ERC-4337 creates a separate smart-contract account and routes actions through bundlers and an EntryPoint, without changing Ethereum’s core rules. EIP-7702, live since the Pectra upgrade in May 2025, upgrades your existing account in place: it keeps the same address and key but lets that account run contract code, and the delegation is revocable.

Are smart-account wallets safe after the Radiant and Bybit hacks?

The contracts themselves were not broken in Radiant, WazirX, or Bybit; in each case attackers tricked human signers into approving malicious transactions that their screens misrepresented. Smart accounts are safe when you can verify what you sign, which means using clear signing, refusing to blind-sign hashes, and, for large sums, a multisig with a disciplined review routine.

Do I still need a seed phrase with a smart account?

Often not. Many smart-account wallets replace the seed phrase with a passkey held in your device’s secure enclave plus social recovery through guardians. That removes the single most common cause of catastrophic loss, though you should still understand and safeguard whatever recovery method your specific wallet uses.

Does the SEC treat a smart-account wallet as a broker or custodian?

A 2026 SEC staff statement said neutral software that only lets you transact from your own self-hosted wallet is not a broker, as long as it does not solicit, route orders, or take custody of your assets. A self-custody smart account you control yourself is not treated as a regulated intermediary, and your tax obligations on gains and losses are the same as for any other wallet.

By Yuki Tanaka, senior wallets and security correspondent at HOGE Wire.