I won't introduce myself too much, since everyone already knows you. I'm Maodi from Wu Blockchain. In this episode, we're interviewing Ben, co-founder and CEO of Bybit. So, first question: you previously said the money stolen by the North Korean hackers was equivalent to about a year's profit for the company at the time. Have you recovered from that incident?
Actually, we probably already covered that hole in terms of profit around last year. So I wouldn't call it a hole — at the time, we used our own company treasury, our own funds, to make the customers whole. But that money was basically earned back through fees by around the end of last year.
OK. Then, if North Korean hackers can mobilize the power of a state to attack exchanges, does that mean it's only a matter of time before other mainstream exchanges get hacked too?
I think, first of all, North Korea has already mobilized state power — the Lazarus group. And recently you've also seen Bitget have some incidents. Actually, since Bybit, it's not just Bitget — hacking incidents across the board are still very frequent. So if you ask me whether other exchanges will get hacked in the future, I think at least for now, there will definitely be new exchanges getting hacked.
And I think the warning for many exchanges should be: don't make your first assumption "I won't get hacked." Your first assumption should always be "I will get hacked," and then under that assumption, figure out how to make the hack survivable — something you can accept, so it doesn't instantly break your capital chain or lose all customer funds. So your entire security architecture needs to spread funds across different places, and then under that assumption — because I think quite a few exchanges, some publicly, some not, some in the customer funds segment, some in their own funds segment — I'd guess quite a few have had hacking incidents.
So I think for many exchanges, most people's assumptions or security plans should now be based on "I will get hacked," "I will be breached by hackers," "the hacker is already inside the house" — and build zero-trust security defenses on that basis. That's how it is.
You just mentioned the Bitget hack. What's your take on this incident? And was it somewhat similar to what happened to you back then, since both involved third-party security software or products having problems?
There are similarities and differences. I think the similarity — well, not really a similarity — is that it was indeed a third party. We looked at their security report: it was a security access-control software, the kind that verifies whether someone is an employee. I didn't look at the specific details in depth; our security team did a full analysis and briefed me. Ultimately, the hackers infiltrated their entire system — you could think of it as their whole infrastructure, including possibly their wallet system — forged upstream withdrawal requests, bypassed the risk-control system, and the hot wallet treated it as a genuine command and processed the withdrawal.
So the final conclusion, from what I saw, was that the hot wallet's keys were not compromised, but the upstream withdrawal information was forged. That's the fundamental difference from us: we were hacked through a third-party wallet provider. And that also explains — because recently I saw CZ give an interview — going back to the old question of why Bybit didn't shut down, didn't suspend withdrawals after being hacked. It's because we knew very clearly that we ourselves hadn't been compromised, that the hacker hadn't gotten inside our house, that our entire system hadn't failed. We had put a lot of Ethereum into an external wallet called Safe, a multi-sig wallet. One of that wallet's main developers was socially engineered and hacked, and code was injected into its web interface, so that when Bybit signed, what we saw was fake. So ultimately, you could think of it as: I put my money in an external vault, and it got stolen. That's why we were able to process withdrawals at the time.
Looking back, if we ourselves had truly been hacked, like Bitget, I think we would have suspended withdrawals too. So I think this is something many people — we have similarities, but also many differences — with this Bitget incident.
Now AI technology is developing quite fast, and many hackers are probably using AI to enhance their attack capabilities. As the defending side, do you think exchanges will face more and more difficulties and troubles in terms of security?
Right. First of all, as an exchange, we definitely have to use AI to scan our own code too. So if others are going to scan us, we certainly need to scan ourselves, and if we find any high-risk bugs or whatever, fix them. And then the question you raised — what AI's emergence brings to exchange security overall — for us, it's not really about some external thing called AI. Rather, I think most exchanges, including us, at least Bybit, are using AI very, how should I put it, extensively or very aggressively. In our internal development and coding, if you look at our pull requests, a lot of our code now is written by AI.
And that includes our internal efficiency gains — not just product and R&D, but even frontline marketing teams, BD teams, and most importantly, customer service and BI teams, data teams — all have started using AI to integrate internal systems. I set a goal for the team: if an employee joins the company and needs to access 10 systems, in theory their internal BybitAI should be able to access those 10 systems, so they don't need to access them directly — they just talk to the AI, and it helps them execute all kinds of operations.
About half a year ago, we equipped every employee with a role-based AI. This role-based AI knows all of that role's work documents, all the meeting minutes, that whole set of materials. So at any time, if the employee forgets — say, "what deal did I make with so-and-so three years ago" — the AI can immediately tell them and remind them. And another great advantage is that when someone in that role leaves, we wipe their personal information, and the new person who comes in can continue to inherit that brain. So there's a role brain, and then partial company-level, even department-level, and whole-company-level brains. So going forward, if we face security issues — and of course I've described a lot of AI usage — the biggest challenge is the security infrastructure for all these AI tools we use. So we now have an AI security department building a whole AI network layer, an AI layer wrapped around us, to protect everything we currently use AI for, ensuring our company's internal secrets, some employee information, and even up to the top-level user information are not leaked. I think these things are all quite important.
You just mentioned a point — I remember Anthropic released a report a while back. The gist of the report was that they could see a lot of users' usage information, and some of it — whether from national security agencies or private companies — that sensitive user information was all being aggregated at Anthropic, the parent company. So when you use these services yourselves, it can't be your own self-developed large model — I understand you're also using external large language models. Are you also worried about your core secrets being leaked?
That's a very good question. I asked our tech team too, and I read that report — it startled me as well. First of all, we work directly with Amazon Web Services to use their services, for example Anthropic, within a specific environment. And when negotiating with AWS, they absolutely will not store any user information — our usage information — they guarantee it's not stored. So we don't directly use Anthropic's or these big vendors' products; instead, we signed a contract with AWS ensuring they absolutely will not store any of our data. How they actually achieve that — whether they have their own models, how they negotiate with Anthropic — some viewers may be more familiar with the technical details. But I did confirm many points with our tech team at the time to look into this.
Also, our approach — we use this setup primarily as the ultimate line of defense. And of course, before that, we definitely encrypt all our information in various ways as well.
I have some friends at exchanges, and in offline chats they mentioned — of course, other exchanges — that the AI products their exchange developed for them require all kinds of authorizations because they may involve sensitive information. So in their own usage, they actually feel it's better to just open their own Anthropic or OpenAI account. Doesn't that bypass what you just described — going through AWS so the conversations won't be leaked? Because they're using it as an individual, but during that usage it's hard to avoid involving some company business information.
You mean employees? Right — specifically for each employee. First of all, if they use their own personal accounts, they can't connect to any of our internal systems, nor can they access any of our internal data. All our gateways — you can think of it as passing through our AI gateway and a whole-company gateway. Any external AI that we haven't authorized cannot connect to any data. And then, for example, our developers — although they use a computer, that computer actually runs in the cloud, meaning their local machine also can't access our data. So we've done a lot of encryption and protection at different layers in this regard. So of course, if an employee wants to use some random tool on their own, that's up to them, but they definitely can't access or reach our internal data.
And our internal systems are further divided into several types — some can directly access our code, and those are heavily isolated from the external network. So it's divided into several different layers. I think we use these methods to handle this.
Just yesterday, OKX's founder Xu Mingxing said that OKX currently spends over 10 million dollars a month on AI. How much does Bybit currently spend on AI per month? In terms of the company's headcount and organization size, you're not in the same league — the number of people is different, right? So the bill for this must not be small either.
But we've been keeping it under control. Haha, I'm not sure if that's a good thing or a bad thing.
Then, after the hack in February 2025, as CEO, are you now devoting more of your attention to security matters?
Yes, definitely — I've put a lot of energy into this, and made many organizational and personnel adjustments. That's one part. Another part was bringing in external, professional third-party security firms to do overall penetration testing for us, as well as code-level penetration testing of all our wallet systems and deposit/withdrawal risk-control systems. And even for our deposits and withdrawals — since we have cold and hot wallets, there are personnel aspects. Beyond internal audits, we also have external auditors reviewing all our logs, deposit/withdrawal records, and videos. For example, for every withdrawal, who initiated the request — all of this has to be documented. And this is also because we've obtained many global licenses — now the EU licenses also have a lot of these requirements — so overall, I think the standard has been upgraded a lot.
Oh, that's at the company level. Personally, do you also spend more energy on this?
