Binance has confirmed that the hackers behind the largest crypto theft on record used one of the exchange's own AI tools to widen their attack. In an interview on TheStreet Roundtable, Mace A
Binance has confirmed that the hackers behind the largest crypto theft on record used one of the exchange's own AI tools to widen their attack.
In an interview on TheStreet Roundtable, Mace AI co-founder Walton Comer gave his analysis of that failure. The AI tools on a compromised company laptop, he said, had more access than the engineer using them.
Mace AI describes itself as an agent assurance company for enterprise AI. It tests AI models and agents the way a hacker would, looking for cognitive weaknesses, and fixing them before real attackers find them. Its product, Fisher, adversarially tests AI agents that use tools and then gives companies the evidence and steps they need to fix what it finds.
Related: Binance says its $6.3 billion hack started with a fake job offer
Binance's initial findings report, published Nov. 3, puts the total impact at about $6.34 billion in digital assets. About $2.85 billion has been frozen or intercepted so far.
The report describes a months-long campaign. It began in June with a tampered software download, moved through fake job interviews with a Binance engineer in September, and escalated on Oct. 20, when the attacker used that engineer's compromised computer and an internal AI tool to reach two more engineers on the wallet infrastructure team.
On Oct. 29, the attacker used stolen credentials to deploy a container and trigger unauthorized withdrawals from Binance hot wallets. Hot wallets are the internet-connected wallets an exchange uses for day-to-day transactions. The FBI has indicated North Korea is responsible, Binance said in the report. The FBI said the same in a public notice on Oct. 31.
Comer co-founded Mace AI, which tests companies' AI agents for security weaknesses. He has dual experience on Wall Street, starting his career with Steve Cohen and later succeeding in working with institutional partners in the high-risk digital assets space. He previously spoke to TheStreet Roundtable alongside co-founder Greg Frank about what OpenAI's Hugging Face breach means for businesses.
He said the attack chain follows a pattern the industry knows well, with one new step. He praised Binance for publishing its findings - "they show real leadership with that" - but stressed that beyond what the exchange revealed, "we don't have the exact details."
His assessment of how the breach widened is blunt. "It would be shocking to me if that machine should have had the ability to have the keys to anything. My gut is a rather typical engineer's laptop."
Related: Mace AI funds AI safety research at University of Virginia’s School of Data Science (UVA)
A laptop like that would normally hold credentials for certain systems, maybe code-commit access, but not wallet keys, he said. But "it had access to AI. And the AI, unfortunately, had access that Binance did not realize the AI had access to."
"Once the North Koreans were on that machine, they were able to talk to the AI that was also on the machine. And that AI is what helped them get access to additional machines within Binance's infrastructure."
"Had they not been able to do that, then it probably would have just stopped right there at that one laptop... That's where it should have stopped."
Comer said he was not surprised Binance was the target. "They are the juiciest target there is for North Korea. That is their number one target," he said. "Binance is by far one of the most accomplished and capable shops there is in terms of cybersecurity and their practices are some of the top in the world." Getting in, he said, "went through a lot of steps... It wasn't trivial. This was a very complex attack."
For Comer, the fix is a rule he says should be absolute. "No agent should be able to do something that the person talking to it does not have permission to do."
He said there are many ways to get an AI to act past its alignment and "one of them is certainly to just ask nicely." Even well-built guardrails tend to fail. "It's almost always the case that the guardrails that are in place don't work as well as people would have assumed. And the only way to find out is through testing."
"If you're monitoring, then you tend to find out after a thing has happened... If you're testing, then you're able to identify where the holes are in advance. And you need both of these, and you need them internally and you need them externally."
Binance's postmortem says the exchange has rebuilt the affected internal AI tool and its access controls, replaced devices, decommissioned credentials and engaged independent security firms for further review.
"If you look at the postmortem and the announcement from Binance, they announce that they're doing both of those things. That's the correct action," he said. "Every company in the world that's deploying agentic tool chains is going to have to start taking these same approaches and they're going to have to be very serious about it. I think this is a wake-up call."
Agentic tool chains, in plain English, are AI agents wired into a company's systems so they can take actions on their own.
Comer warned that the Binance breach is not the end of it. "October was pretty bad when it came to cybersecurity," he said, describing a stretch people are calling Black October. "November's probably gonna be worse. And the bad news is December is gonna be even harder."
Comer drew a line between exchange security and Bitcoin itself. "I don't think Bitcoin is nearly as at risk from AI," he said. He said Bitcoin has already been heavily scrutinized by AI models, including models from Anthropic and OpenAI that are not public, "and so far nothing has been found that is ... a deal breaker."
The bigger long-term risk, he said, is quantum computing rather than AI. "The risks are more to do with cryptography and quantum computing than anything else," he said, adding that both Bitcoin and Ethereum already have plans to adapt if quantum attacks become real.
He closed with a warning that there is "a very high likelihood that we're gonna be talking about the next one of these major headlines again."
