When Ethereum's Pectra upgrade shipped EIP-7702 on May 7, 2025, the promise was simple: let ordinary externally owned accounts (EOAs) act like smart contracts without changing addresses. The reality, three months later, is a security nightmare. A new study by USENIX researchers, set to be presented in 2026, has analyzed 22.8 billion historical transactions and found that 63% of the 3.66 million EIP-7702 delegation transactions executed so far are malicious. The exposed value stands at over $2.36 million in direct theft, with another $10.14 million at high risk of being drained. The attack surface is not a bug in a single contract—it is a systemic flaw in how Ethereum's most fundamental account model now trusts code.
The Architecture of Betrayal
EIP-7702 was designed to be a pragmatic step toward full account abstraction. Instead of forcing users to migrate to new smart contract wallets, it allows an EOA to sign a one-time authorization that delegates its execution logic to a designated contract code. That code can then run any logic—token swaps, DeFi interactions, NFT mints—on behalf of the EOA, while the EOA retains its original address. The upgrade was hailed as a low-friction path to give millions of users smart wallet capabilities without ecosystem fragmentation.
But the USENIX study reveals the fault line: the delegation is not just a UI convenience; it fundamentally rewires the trust assumptions of the Ethereum Virtual Machine. The check msg.sender == tx.origin, long relied upon by dozens of DeFi protocols and front-end protections to prevent contract-based phishing, becomes unreliable when an EOA itself is executing code that can be replaced at any time. The attacker's code can impersonate the user, bypass signature checks, and drain allowances—all while the user's wallet interface shows a benign "delegated" status.
The researchers dissected 242 unique malicious contracts that have already been deployed. They fall into two distinct categories. The first, "proto-scam" contracts, mimic legitimate DeFi protocols and trick users into delegating to them via fake airdrops or phishing sites. The second, more insidious class are "re-binding" contracts. These allow an attacker to first obtain a delegation, then quietly swap the delegated code to a completely different contract that the user never signed. Because the delegation is keyed to the EOA and not a specific code hash, the user's wallet may still show a green "connected" indicator while the attacker's new code executes arbitrary logic. In effect, the user has unknowingly signed a blank check.
The CREATE2 Pre-commit Trap
A particularly alarming finding is the use of 500 pre-computed contract addresses via CREATE2. Attackers can deploy contracts at predetermined addresses, wait for users to delegate to those addresses, and then later activate malicious code. Because the address is fixed, it can be whitelisted by a wallet or a dApp before the code is even live. The researchers call this "delegation to a ghost"—a promise of code that does not yet exist, but which will turn hostile on demand. This technique is already being used to bypass security warnings that rely on contract verification status or code existence checks.
The Liquidity Impact
The direct financial losses are small relative to Ethereum's total value locked, but the systemic risk is outsized. The study estimates that 24% of ETH and ERC-20 tokens held in wallets that have interacted with 7702-capable interfaces could be exposed to this attack vector. That's not a single protocol's vulnerability; it's a contagion risk across the entire L1 ecosystem. The problem is compounded by the fact that many DeFi protocols still use tx.origin as a coarse anti-bot or anti-phishing measure. With EIP-7702, those checks are now trivially bypassed, leaving protocols like old versions of MakerDAO, some bridging contracts, and even certain NFT marketplaces open to instant asset theft.
Market reaction has been muted so far, but the implication is clear: the narrative of account abstraction as a pure upgrade is broken. "We are seeing a structural break in the security model of Ethereum," said one researcher who reviewed the study. "The delegation primitive is a powerful tool, but it was shipped with a trust model that assumes the user will only delegate to audited, immutable code. The reality is that over 60% of delegations are to malicious actors, and the infrastructure to differentiate good from bad code is not in place."
The Defense Gap
Why hasn't the community reacted? One reason is that the attack footprint is highly fragmented. The 3.66 million malicious delegations are spread across thousands of addresses, with average losses of only a few hundred dollars per victim. This is not a single high-profile hack that triggers a market-wide panic. Instead, it is a slow bleed—a "silent deleveraging" of user trust. The researchers warn that the situation could escalate rapidly if a large DeFi protocol is exploited via a re-binding attack, or if a widely used wallet fails to implement proper delegation warnings.
The report calls for immediate action on three fronts. First, wallets must implement delegation whitelists that check code hash integrity rather than just address matching. Second, protocols that rely on tx.origin must be upgraded to use msg.sender combined with permit-based signatures. Third, the Ethereum community should consider a supplementary EIP that introduces a delegation lock-in mechanism, where a user can commit to a specific code hash and prevent re-binding without an additional signature. The latter is technically feasible but would require another hard fork, a process that could take 12–18 months.
The Account Abstraction Cold War
The findings also cast a shadow over the broader account abstraction roadmap. ERC-4337, which operates via a separate mempool and does not alter the EOA semantics, had been seen as the safer but slower path. EIP-7702 was the aggressive shortcut. Now, the security community is questioning whether the shortcut was worth the cost. "This is the classic tension between protocol evolution and adversarial adaptation," says the research team. "The attackers are always faster to adopt new primitives than the defenders, because the defenders need to coordinate across wallets, dApps, and standards bodies. EIP-7702 is a textbook case of this asymmetry."
In the meantime, the researchers maintain a public blacklist of the 242 malicious contracts and the 500 CREATE2 pre-commit addresses. They urge wallet developers to integrate this list immediately. But they also caution that the list is reactive; new attack contracts are being deployed every day, often using automated scripts that generate fresh phishing sites and delegation requests. The true fix is a redesign of the delegation semantics, and that will take time.
For the average user, the advice is stark: unless you fully understand the code you are delegating to, and have verified that the code cannot be re-bound, treat any EIP-7702 delegation prompt as a potential total loss of the associated wallet's assets. The narrative of "self-custody with smart features" has collided with the hard reality of permissionless adversarial code, and for now, the code is winning.
