A maze of wallets: what it takes for AI to follow the Bitget millions

On 24 September, attackers took about $387 million from Bitget's hot wallets across seven blockchains. Bitget's CEO has attributed the attack to North Korean groups, and the exchange is offering a 5% bounty for help recovering or freezing the funds.

Since then, the money has been swapped into Bitcoin and spread across thousands of new wallets. Investigators following it are leaning on AI agents because no team can watch that many wallets by hand. But an AI agent can only take them so far. To turn a trail into a frozen account, it has to know who is behind each address it reaches.

Thousands of wallets, one investigator

We spoke to an investigator we work with who has followed the Bitget funds since day one.

The scale is the problem. "The complexity of this case is the number of wallets," he says. "You cannot do it manually." The attackers also send funds back and forth between wallets to break the trail, so an alert on a single address goes quiet after one hop.

His answer is an AI agent of his own, which follows the funds across all of those wallets at once. Which model it runs on matters less, he says, than the data it can reach.

Why the money is sitting still

For now, most of the Bitcoin is idle, and in the first weeks the attackers stayed away from exchanges altogether. They cannot move this much at once without being noticed, and the sum is too large to swap into a privacy coin such as Monero, which does not have the liquidity to absorb it. So the funds are expected to move slowly, in smaller pieces, through OTC desks, mixers and regional exchanges.

That is where this kind of case is won or lost. Exchanges and OTC desks hold customer accounts. When stolen funds reach one, there is a business that can freeze them and knows who the customer is, but only if the investigator learns quickly that the address belongs to one.

One question at every hop

Each time a wallet moves, the AI agent reaches a new address and has to decide what to do with it. The investigator's rules come down to one question: who is this?

• An exchange or OTC desk: stop and flag it, so the investigator can ask that business to freeze the funds and say who holds the account.
• A mixer: note where the funds went in, and watch for what comes out.
• Another of the attacker's own wallets: skip it, instead of going round in circles.
• Unknown: keep tracing.

Without the context of labelled wallets, the AI agent is blind. "Agents are capable of following the funds," he says. "The problem is they don't know what the address is." An AI agent that does not recognise an exchange keeps going straight through the exchange's own wallets, and ends up following thousands of unrelated customers instead of the stolen funds.

"If you put trash information in, you get trash out," he says. "With a good feed of labels, it is the next level of tracing for the agent."

Why free labels are not enough for a case like this

Free wallet labels are useful for a first look, and the big exchanges are well covered. The gaps sit exactly where the Bitget funds are expected to go: regional exchanges and OTC desks, which few public tools identify. This is where Chainlabs leads the industry: we cover more than 2,000 OTC desks and 800 regional exchanges, many identified first-hand by investigators working in over 50 countries. A miss there is not a small error. It is the freeze that never gets requested.

In a case of this size, an attribution also has to hold up. An exchange's compliance team or a regulator will ask where it came from, and a crowd-sourced label with no source cannot answer that.

Where Chainlabs fits

Chainlabs identifies who is behind crypto addresses. Our network of investigators establishes many attributions first-hand, rather than inferring them from on-chain patterns alone. Our data already sits behind leading compliance platforms, and our coverage is strongest where this case is heading: regional exchanges, OTC desks and darknet services.

Every answer tells you what it rests on:

• whether we hold our own record for that exact address, or inferred it by clustering, and how confident that inference is
• the date of the data, which is updated daily
• whether evidence is available on request, which it is where we traced the address ourselves

Now, inside your own AI agent

So far, our data has been available mainly through the largest compliance platforms. We are now making it available to investigators and builders directly, through the Chainlabs MCP server, starting with a small group of design partners. MCP (Model Context Protocol) is the standard way to connect an AI agent to a data source, so there is no new software to learn. You add the Chainlabs MCP server to Claude, or any AI agent that supports MCP, and it can ask who is behind an address, or check a whole batch, at every hop of a trace. It starts with Bitcoin, where we hold attribution for more than 320 million addresses.

Become a design partner: If you trace funds with AI, or you are building an investigation tool and need attribution you can stand behind, we would like you to test it for free and tell us what is missing.

👉 Email mcp@chainlabs.ai

---

FAQ

Who is Chainlabs?

Chainlabs identifies who is behind cryptocurrency addresses. Investigators and compliance teams use its attribution data, and it sits behind several leading compliance platforms.

What is the Chainlabs MCP server?

It connects Chainlabs attribution data to AI agents through the Model Context Protocol. An AI agent can ask who is behind a Bitcoin address, or check a batch of addresses, and get the entity, its category, what the answer rests on and the date of the data.

Which AI tools does it work with?

Any AI agent or client that supports MCP, including Claude.

How is Chainlabs data different from free wallet labels?

Every answer says what it rests on: our own record for that address, or clustering with a confidence level. Evidence is available on request where we traced the address ourselves. Coverage is strongest for regional exchanges, OTC desks and darknet services, which free tools rarely identify.

What happens when Chainlabs does not know an address?

It says so, with the date of the data and our overall coverage. It never guesses, and never calls an address clean.

Which blockchains does the Chainlabs MCP server cover?

It starts with Bitcoin. Chainlabs covers more than 30 blockchains, so tell us which one you would most like added next.

How up to date is the data?

It is updated daily.

How can I become a design partner?

Email mcp@chainlabs.ai. Testing starts with Bitcoin and is free during the test period.