binance's $6.3b hack started with a fake job interview

daily ai × crypto news in 5 minutes. no hype, just what matters.

evening. yesterday we showed that the laundering of the binance loot runs on scripts, not ai. today binance published its own incident report on how the thieves got in, and this time ai is in the story: not as a hacker super-brain, but as binance's own internal tool, used as a hallway between computers. the door itself was a fake job interview. also tonight: akash votes to leave its own chain for solana (careful with fake tokens), anthropic tells ipo investors it may soon hit its "ai builds ai" line, and a second exchange loses ~$38m to plain signed transfers.

what binance said. in an initial incident report published tuesday, binance said attackers used malware, stolen credentials and "compromised internal ai tooling" to reach its wallet-signing systems and take more than $6.3b on oct 29, the largest crypto heist on record. the fbi attributes it to north korea, which denies it. co-ceo richard teng: "this wasn't one mistake or one breach — it was a months-long campaign." (the block, reuters via nbc)

step 1, june. a wallet devops engineer installed a tampered homebrew package while trying to install an ai coding tool. binance caught it, replaced the device and rotated credentials, but the attackers had already learned his role and who his colleagues on the wallet team were. (the block, coindesk)

step 2, september. a second engineer was contacted on a legitimate job platform by a fake "project director." a fake hr person interviewed him on video and even shared a social media account with personal photos to build trust; a second interview was with a purported ceo. the last step was a coding assignment to download onto his work computer. it carried zero-day malware that endpoint security didn't catch. this is classic social engineering: no code was broken, a person was. (coindesk)

step 3, oct 20: the ai part. from that infected laptop, the attackers used binance's internal ai tool to take control of two wallet engineers' devices, and on one of them found a container-deployment permission. that's the twist: the ai that mattered wasn't an attacker's model, it was the company's own assistant, which had enough reach across machines to be useful to whoever held the keyboard. (the block, reuters via nbc)

step 4, oct 29. nine days later, stolen credentials were used to deploy a container and run a malicious script. funds started leaving at 20:05 utc. the attacker used an ethereum feature, eip-7702, to get execution rights over a hot wallet, outside binance's normal withdrawal flow. because those wallets use multiparty computation (mpc, a cousin of the multisig idea), no single key could revoke it, and containment took until 00:10 utc on oct 30, about four hours. (coindesk)

the damage control. binance says about $2.85b has been frozen or intercepted and about 98% of affected tokens were replenished with company funds, so users won't bear the loss. it says it blocked 57 suspected north korea-linked attempts this year, many with the same playbook of fake interviews and malicious developer tools, and is commissioning an independent review. (the block, coindesk)

why it matters. the security story of the year isn't a broken smart contract, it's a broken hiring funnel plus an over-privileged internal ai. every company rolling out an internal ai assistant with access to employee machines just got a free threat model: if one laptop falls, what can the assistant reach? for you, the lesson is simpler: never run a "take-home test" on a computer that touches money or work credentials. yesterday's #020 showed the getaway was scripts; today shows the door. for a smaller cousin, see how the "steal my bitcoin" server fell in #013. more in crypto.

akash votes 97% to move its ai compute market to solana, but nothing moves yet. akash, a depin marketplace that rents out gpus and servers, passed prop 346 on nov 2: about 59.98m akt voted yes and ~796k no, roughly 97.3% of the ~61.66m akt cast. it approves a 1:1 spl version of akt on solana through a one-way portal ("there is no path back"), expected 2–3 weeks after the vote and after two independent audits. (solana compass, akash)

the marketplace itself needs a provider testnet of at least six weeks with 25+ operators, a 14-day rehearsal and a separate cutover vote; the old chain halts 90 days after that. until then, akt keeps earning staking rewards on the old chain. the warning: tokens with the akt ticker were already created on solana in october, before any official one existed. none of them is real; take the mint address only from akash's channels. (solana compass) related: another depin network paying for ai data in our filming-places case.

anthropic tells ipo investors it may soon hit its own "ai builds ai" line. in its amended s-1 filed monday, anthropic added risk factors saying it has in the past not recognized big jumps in its own development speed until after they happened, and that its internal estimates for future models suggest it "may imminently meet" its responsible scaling policy definition of "dramatically accelerated progress attributable to ai." no current model is designated as past that threshold. (techcrunch)

it also says it doesn't expect to let claude "fully autonomously design and build" successor models without meaningful human supervision and is considering policy updates to keep humans in control. read it for what it is: a legally required risk disclosure, written by lawyers to protect against lawsuits, not a product launch, but also not a nothing. the crypto link: the same coding-capable models are what security teams (and attackers) now point at wallets.

brazil's mercado bitcoin drained across 10 chains, with plain transfers. on nov 2, hot wallets on at least 10 blockchains were emptied; a report relayed by opsek founder pablo sabbatella put most outflows between 06:19 and 06:53 utc, as plain signed transactions, not a smart contract exploit. how the attacker got signing access is unknown, and there's no public link to the binance hack. (coindesk)

the exchange says about $38m was lost, $11m has been frozen with help from tether, circle and chainalysis, and it will cover all user losses. same pattern as binance: when the keys fall, the only fast brake is a stablecoin issuer's freeze button. (coindesk)

aikido safe chain: a free, open-source guard that sits in front of npm, yarn, pnpm, bun, pip, uv and poetry and blocks known malicious packages before they install, plus holds back brand-new releases (default 48 hours) until they're checked. free, no account or token needed. for developers and crypto builders who install packages on machines that touch keys, i.e. exactly the binance attack surface. it won't stop a malicious repo you run by hand, so the "no homework on the work laptop" rule still applies. (aikido) more picks in the gweiwire tools catalog.

binance's full report and the independent review: which internal ai tool it was and what permissions it had. akash's official spl akt mint address and audit reports. and whether mercado bitcoin says how signing access was lost.

that's the wire. the hack didn't break crypto, it broke a job interview. back tomorrow. — gweiwire

forward this to the friend doing a "quick coding test" for a dream job. new here? subscribe at gweiwire.com. not financial advice. last sunday's case: how to earn crypto filming places for ai.