Binance co-CEO Richard Teng said the crypto industry must learn from the $6.3 billion Binance hack, which involved a new Ethereum feature, fake recruiters, and AI techniques. The attack has raised questions about whether new crypto tools increase risk.
Teng's account of the attack
Teng told CNBC on Tuesday that Binance published a detailed report so the industry can learn from the incident. He cited an FBI attribution to North Korea and described a campaign built over months.
According to Teng, the campaign began with a fake developer tool in June, followed by fake recruiter interviews in September. Attackers then moved through internal AI tools to the system that authorized wallet transactions.
On October 29, they exploited a new Ethereum feature that kept funds moving after Binance paused withdrawals, Teng said. On Ethereum, one wallet lost about 167,464 ETH and $1.6 billion in Tether (USDT).
The Ethereum feature at the center
Researchers identify the feature as EIP-7702, an Ethereum upgrade from May 2025 that lets ordinary wallets borrow smart-contract code, according to Gizmodo. Blockchain security firm Blockaid found that the attacker bundled 50 signed authorizations into one transaction. Binance cleared the delegation nearly four hours later, per a detailed account of how the hack happened.
Ethlabs researcher Derek Chiang argued that EIP-7702 is not broken, though it made the theft faster and more efficient. Holders using wallets that support the feature, such as MetaMask, need not panic, he added.
AI tools and custody design
AI tools also featured in Teng's account, both as attacker techniques and as an internal route to the signing system.
Binance says it has replenished 98% of affected assets, rebuilt the wallet system, replaced devices, and cancelled linked credentials. It is also working with law enforcement and industry partners to trace and freeze stolen funds.
If the independent review Teng promised confirms signer compromise, risk sits with exchange custody design rather than new features alone. A flaw in a feature itself would change that read.
