Hook
Three million six hundred sixty thousand transactions. Sixty-three percent malicious. That is not a typo, and it is not a slow-motion threat. It is the verified on-chain record of EIP-7702's first three months on Ethereum mainnet, measured by USENIX security researchers who analyzed 22.8 billion historical transactions to isolate the damage.
The data shows a simple truth: attackers reached the new account abstraction paradigm faster than the defense community could build guardrails. The upgrade activated on May 7, 2025, as part of Pectra. By August, the attack surface was not theoretical. It was operational, automated, and profitable.
Ledgers do not lie, only the narrative does. The narrative said EIP-7702 would make Ethereum wallets smarter. The ledger says it made them vulnerable.
Context
EIP-7702 is a consensus-layer change to Ethereum that allows externally owned accounts (EOAs) — the ordinary wallets users have held for years — to temporarily delegate their authority to smart contract code. The address remains constant. The private key remains in user custody. But during a transaction, the EOA can execute arbitrary logic as if it were a contract.
This is the long-awaited evolution of account abstraction. ERC-4337 introduced a parallel system of smart contract wallets operating in a secondary mempool. EIP-7702 goes further: it changes the base semantics of the EOA itself. No migration needed. No new address. No asset movement.
The design is elegant. The risk is asymmetric.
Under the old model, a private key was the sole authority. Lose it, lose everything. Attackers targeted key theft. Under the new model, a private key still matters, but it now co-exists with delegated code. An EOA can authorize a contract to act on its behalf. The key remains the root of trust. The delegation, however, introduces a second attack vector that did not exist before.
Traditional EOA security was binary: the key is compromised or it is not. EIP-7702 creates a gray zone. The key can be safe. The delegated code can be malicious. The user can sign a blind transaction, delegating authority to a contract that drains assets in one move. This is not a hypothetical. The on-chain record confirms it at scale.
Core
The USENIX team performed a forensic sweep of 22.8 billion historical transactions, identifying 3.66 million EIP-7702 delegation transactions in the first three months following activation. Of these, 63 percent were malicious. Let that ratio settle.
Sixty-three percent of all delegation activity was not legitimate adoption. It was attack traffic.
The attack patterns break down into two dominant categories. First, protocol-related phishing: malicious contracts that impersonate legitimate services, tricking users into signing delegations that transfer control. Second, deceptive re-binding: attacker contracts that rebind an EOA's delegation after a legitimate benign delegation has been made, silently replacing trusted code with malicious code.
The re-binding vector is the more dangerous finding. It defeats the user's expectations. A wallet holder believes their address is delegated to a reviewed, legitimate smart contract. In reality, the delegation has been overwritten to point at attacker-controlled code. The wallet interface may still display "normal" status. The scan may not flag the change. The device shows a healthy wallet. The assets are gone.
This is not a peripheral vulnerability. It breaks the fundamental defense mechanism that DeFi protocols have relied on for years: the assumption that msg.sender and tx.origin are reliable indicators of user intent. When an EOA is delegated to malicious code, msg.sender no longer represents the human. It represents the delegated contract logic. Any protocol still using tx.origin as a phishing guard is now exposed to a bypass that did not exist before Pectra.
The researchers also identified 242 known malicious contracts and an additional 500 contracts deployed via CREATE2 that remain unverified. Code is law, but bugs are inevitable — and when the law is unwritten, it is unenforceable. These 500 unverified deployment addresses constitute a dormant threat. They have not acted. They can.
The financial damage is measurable: $2.36 million directly attributed to these attacks, with an additional $10.14 million exposed to potential theft through vulnerable delegation configurations. Compared to the scale of ETH itself, these figures are small. Compared to the trust architecture of account abstraction, they are a warning flare.
Contrarian
The natural conclusion is that EIP-7702 is too dangerous and should be restricted. That conclusion is wrong. The data does not support abandonment. It supports a maturation pause.
Consider the alternative reading: 37 percent of delegation transactions were benign. The technology attracted real, organic usage within weeks of activation. That is adoption velocity. The problem is not that the mechanism is broken. The problem is that the defense ecosystem around it has not caught up.
Volatility reveals character, not just value. The same principle applies to infrastructure. An upgrade that works flawlessly under benign conditions but collapses under adversarial pressure lacks resilience. EIP-7702 has not collapsed. But its early record shows that attack automation scales faster than security tooling.
The correlation between adoption volume and malicious volume is also worth interrogating. High transaction counts do not equal high-quality adoption. A significant portion of the 3.66 million transactions may be attacker-initiated in the first place — bots testing delegation frameworks, enumerating vulnerable wallets, automating phishing campaigns. The 63 percent malicious figure is not a proportion of innocent users being attacked. It is the share of total activity that was hostile to begin with. That distinction matters. It means the legitimate user base is smaller than the raw transaction count suggests, and the per-user risk is actually higher.
Resilience is built in the red, not the green. The current data set is early red. It is instructive. It is not terminal.
Takeaway
Survival is the ultimate alpha in a bear, and scrutiny is the ultimate defense in a bull. The market is treating EIP-7702 as a settled feature. The on-chain record says it is an open question.
The signal to monitor is the malicious transaction ratio. If it crosses 70 percent, the attack model has outpaced the defense model and broader account abstraction confidence will suffer. If it declines below 50 percent, the ecosystem is learning. Additionally, watch for wallet providers shipping explicit delegation whitelists and zero-trust authorization APIs. That infrastructure does not meaningfully exist today. It must exist within the next twelve months.
Do not delegate to code you have not audited. Do not rely on tx.origin for security. And when a research report presents a 63 percent attack rate, do not look away. Trust the math, ignore the hype.
The ledger has spoken. The next question is whether the ecosystem is listening.