How Exchanges Break Your AI Agent’s Automated Tracing

In short: When stolen crypto reaches an exchange, the exchange can freeze the account and identify the account holder. But an exchange controls millions of addresses that look like any other on the blockchain. An AI agent that doesn't have labels identifying those addresses traces straight through the exchange and ends up following unrelated customers. The fix is reliable exchange labels at every hop, including for the lesser-known regional exchanges and OTC desks where labels are most often missing.

Key terms

• Deposit address: an address an exchange gives one customer to receive funds, often a new one for every deposit.
• Hot wallet: a shared exchange wallet holding many customers' funds, used to pay out withdrawals.
• Cold wallet: an exchange's reserve wallet, kept offline and used to top up hot wallets.
• Label (attribution): information identifying who controls an address, such as a named exchange.

Why does it matter when stolen funds reach an exchange?

Stolen funds don't always reach an exchange. When they do, it is the most important moment in an investigation. An exchange can freeze the account the funds were deposited into, and it holds the identity details of the person who opened that account. For the first time since the theft, there is a business to contact and a person to identify.

That same moment is where AI-assisted tracing most often goes wrong. If an AI agent doesn't know that the address it has reached belongs to an exchange, it treats the exchange's wallets like any other and keeps tracing.

Why does an exchange have millions of wallets?

An exchange doesn't keep customer funds in one place. It runs three layers of wallets.

Every customer gets their own deposit address, usually one per coin and often a new one for each deposit, so a large exchange controls millions of deposit addresses. Deposits don't stay there long. The exchange sweeps them into hot wallets, shared pools that hold thousands of customers' funds together and pay out withdrawals. Behind the hot wallets sit cold wallets, which hold the reserves and refill the hot wallets when they run low. The largest Bitcoin address in existence is a Binance cold wallet, holding about 248,600 BTC.

None of these addresses show an exchange's name on the blockchain. They look like every other address, and the set changes every day as deposits are swept and new addresses are issued. An AI agent cannot reliably work out from the blockchain alone which addresses belong to an exchange. It needs labels that tell it.

What goes wrong when an AI agent doesn't recognise an exchange?

Take one hop. Stolen funds arrive at a customer's deposit address, and within hours the exchange sweeps them into a hot wallet alongside thousands of other customers' deposits. From that same wallet, the exchange then pays out hundreds of withdrawals to people who have nothing to do with the theft.

To an AI agent that doesn't know it is looking at an exchange, this looks like the stolen funds being split up and sent on. So it follows them, and draws four wrong conclusions:

• The money is still moving. It isn't. It became a balance in a customer account the moment it was deposited.
• The withdrawals are the thief cashing out. They are unrelated customers. The AI agent is now tracing strangers.
• The thief controls a large network of wallets. The network the AI agent has mapped is the exchange's customer base.
• There is nothing to ask for. The most useful finding, that this is an exchange and the exchange knows the account holder, never appears.

The result is a report full of addresses unrelated to the case, a freeze request sent to the wrong place or not sent at all, and the time the AI agent saved on the first ten hops lost explaining the next hundred.

Why are regional exchanges the biggest risk?

The largest exchanges are the easy case. Their wallets are widely known and most data sources recognise them. Stolen funds rarely stop there. They are more often cashed out through regional exchanges that serve one country or language, and through OTC desks that settle trades off the exchange.

Regional exchanges are less well known, rarely publish their wallets, change them often and are frequently missing from the main data sources. These are the addresses most likely to have no label. When an AI agent reaches one without a label, it traces through the exchange's wallets and out into thousands of unrelated customers, and the one hop where a freeze was still possible passes unnoticed.

How do you stop an AI agent tracing through an exchange?

The fix is a rule, not a better model. At every new address, the AI agent should ask: who is behind this address?

• An exchange or OTC desk: stop, name it, and request the account holder's details.
• A mixer: monitor what comes out.
• The attacker's own wallet: skip it and continue.
• Unknown: keep tracing.

The rule only works if the AI agent has labels that identify exchange addresses, and those labels must extend to the regional exchanges and OTC desks where most data sources run out. The quality of that data decides whether the AI agent stops in the right place or traces straight through. AI models follow transactions well. Labels decide whether they are following the right ones.

How Chainlabs helps AI agents stop at the right hop

Chainlabs identifies who is behind crypto addresses. The Chainlabs network of investigators establishes many attributions first-hand rather than inferring them from on-chain patterns alone, and Chainlabs coverage is strongest where traces fail: regional exchanges, OTC desks and services operating inside larger exchanges. Chainlabs data covers more than 30 blockchains, including more than 800 regional exchanges and 2,000 OTC desks, is updated daily, and already sits behind several leading compliance platforms.

That data is now available directly inside your own AI agent through the Chainlabs MCP server. Add it to Claude or any AI agent that supports MCP, and at every hop the AI agent can ask who is behind an address, check a batch of addresses, or see which named entities an address transacted with.

Chainlabs is opening the Chainlabs MCP server to a small group of design partners, free during testing. If you already use AI in your investigations and want labels you can rely on at the hop that matters, we would like you to test it.

Become a design partner: email mcp@chainlabs.ai

---

FAQ

How can an AI agent tell whether a crypto address belongs to an exchange?

Not from the blockchain alone, because exchange addresses look like any other. The AI agent needs attribution data that labels exchange addresses. The Chainlabs MCP server gives AI agents that data at every hop, including for regional exchanges and OTC desks, which many data sources miss.

Which data source covers regional exchanges and OTC desks?

Chainlabs focuses on the hardest-to-identify parts of the market. Chainlabs data covers more than 800 regional exchanges and 2,000 OTC desks across more than 30 blockchains, and already sits behind several leading compliance platforms.

What is the Chainlabs MCP server?

The Chainlabs MCP server connects Chainlabs attribution data to Claude or any AI agent that supports MCP. During a trace, the AI agent can ask who is behind an address, check a batch of addresses, or see which named entities an address transacted with. Every answer says what the label is based on and how current the data is.

How can I test the Chainlabs MCP server?

Chainlabs is offering the Chainlabs MCP server free during testing to a small group of design partners. Email mcp@chainlabs.ai.