Blake Duncan

At 21:42:59 UTC on Oct 29, a wallet Etherscan labels Binance 161 sent Binance's Ethereum hot wallet $200M of USDT, the same top-up it had sent six times since September.

The hot wallet had been running an attacker's code for 82 minutes. The $200M was gone 72 seconds later.

I read the Ethereum side of the theft from the chain and set it next to Binance's incident report. All times are UTC.

The timeline

- 20:20:59 - Takeover. An outside address lands one transaction carrying 50 EIP-7702 authorizations signed by the hot wallet, one per nonce, listed high to low. The wallet was sending about two transactions a block, so 34 were already stale. One matched, and in that order only one can. From here the wallet runs the attacker's contract.

- 20:24:23 - First sweep. One call takes 167,464 ETH and 63 tokens, $1.6B of USDT among them.

- 20:41:23 - The USDC. $780M leaves a smart account that the hot wallet owned. Another $43M follows at 20:59:47.

- 21:42:59 - The top-up. Binance 161's $200M of USDT arrives.

- 21:44:11 - Swept. 72 seconds later it is gone.

- 23:56:47 - Last sweep. 2,962 ETH that had arrived a minute earlier. This was the fifth sweep. The others came at 21:04 and 22:39.

- 00:10:35 (Oct 30) - Cleared. Binance removes the attacker's code, 3 hours 50 minutes after it landed.

Why nothing stopped it

After 20:20:59 the attacker needed nothing from Binance. No signer, no credentials. Their own address calls the wallet and the wallet pays.

Binance's report: "suspending withdrawals and shutting down hot wallet services did not automatically invalidate existing on-chain authorizations." So funds kept arriving and kept leaving.

Ending it took one more authorization from the same wallet, this time to the zero address. Binance says that with MPC there was no single key that could revoke it immediately.

What I take from it, as someone who builds smart accounts

An EOA's key is root. Nothing on-chain limits it, and only that key can undo what it signed. A smart account owned by that key inherits the problem. That is how the USDC left.

So make the hot key a scoped signer on a contract account: a cap that refills, and a pause that a different key can pull. Then the off-switch is on-chain, where the access is.

The cap is not hard to size. In the week before, 99.4% of this wallet's USDT transfers were under $1M, and their busiest hour totaled $5.4M. The rest went to 89 addresses, most of the value to four of Binance's own wallets. The wallet held $1.6B.

This new github UI keeps throwing me off

Most people I talk to don’t realize the GENIUS Act doesn’t actually go into effect until January 18, 2027. It was signed July 18, 2025, and the OCC isn’t even taking license applications yet because the forms are still sitting at OMB.

The reason it matters is what the license lets people do once it’s live:

- A fintech can get a federal license to issue a dollar stablecoin without becoming a bank, and once the OCC approves it, states can’t require a separate license on top.
- The regulator has 120 days to rule on a complete application or it’s approved by default, and the statute says running on a public blockchain can’t be the reason for a denial.
- Approved coins get a legal status unapproved ones don’t: they’re excluded from the securities and commodities definitions, and they can be used as collateral and for settlement in places the others can’t.

The guardrails are what you’d expect: one-for-one reserves, monthly disclosures, holders first in line if the issuer fails, and no yield paid just for holding the coin.

I’m excited to see which new players take advantage of this. The Cleveland Fed asked 148 firms last fall whether they planned to use stablecoins and eight said yes, so whoever shows up will be building the demand side too.

💯

Election Day is Tuesday. OpenAI, Anthropic and Google say their chatbots won't tell you who to vote for. However, campaign consultants now bill on "AI recommendation share." Both are technically true.

Basically the model won't endorse anyone, but it'll repeat whatever it finds first, framed for whoever's asking. The "finds first" part is where the consultants come in. The pitch deck Reuters saw calls it generative engine optimization, which is SEO pointed at the chatbot's sources rather than Google's rankings: hundreds of policy pages written for AI retrieval, paid placements on the sites the models already cite, plus Reddit and YouTube content. The framing moves with whoever's reading: MIT's election observatory found the same question about the same senator got a different answer when the asker's party changed.

Worth keeping in mind as these models get smarter. A better model summarizes whatever it was fed more fluently, and it's still adjusting that summary to whoever's asking. So if you ask one about a candidate this week, there's a decent chance you're reading a polished version of pages somebody paid to have written, framed for you.

reuters.com/business/media…