Binance has given its first detailed account of how attackers took $6.3 billion from the company — and the way in began with an engineer applying for a job. In initial findings published on Tuesday, the world's largest crypto exchange describes a "sophisticated, multi-stage campaign" that was set up months in advance.
The first move was social engineering. A Binance engineer was approached through a legitimate job-hunting platform and invited to bogus interviews with people posing as a company's representatives. The attackers built trust with what Binance called "more advanced and deceptive" methods, including social media accounts carrying personal photographs. At the second interview the engineer was sent an "assignment" to download. It was malware, its code wrapped in obfuscation that got it past detection.
From that machine, the attackers reached a compromised internal AI tool in October and used it to take control of two other Binance engineers' laptops. From there they reached the tools used to drive the exchange's wallet infrastructure, and drained it. Binance called the incident "serious and highly sophisticated" and apologised: "this should not have happened."
The consequences landed on users immediately. Binance suspended withdrawals worldwide until the service was restored on Sunday. The company says it has set aside more than 100 percent of the funds needed to cover the loss, with 98 percent of the affected token assets replenished as of Saturday. The FBI and Secret Service are assisting with tracing, and Binance now cites the US agencies' attribution of the attack to North Korea. Pyongyang denies it; Kim Yo Jong called the accusation a "sinister attempt" by Washington to damage the country's image.
What Binance says it will do next is the part worth reading closely: invest more in AI tools that detect security threats, bring in third-party firms for a full-scale review, and set up a security research fund. The tool the attackers used as a stepping stone is the one it intends to expand — on the defending side.
Why it matters: Internal AI assistants are now wired into company data, ticketing, code and infrastructure access. That reach is precisely what makes them a worthwhile intermediate target: compromise an agent and you inherit its permissions without needing a second exploit. Binance is the most expensive demonstration so far that an internal AI tool's permission boundary has to be audited like a production database's.
What to watch: whether other exchanges and financial platforms publish comparable postmortems, and whether "we run an internal AI tool" starts appearing in security reports as an attack surface in its own right rather than an efficiency measure.
