1. What Elliptic found about the Binance laundering
Blockchain analytics firm Elliptic said on Monday, November 2, that it found no evidence that artificial intelligence is directing the laundering of the roughly $6.4 billion stolen from Binance on October 29. In its assessment, North Korean hackers spread the proceeds across about 20 automated software services in less than three days, and the pattern looks like human-supervised scripts rather than autonomous AI agents.
The hackers moved funds through decentralized exchanges, lending protocols and cross-chain bridges. Each of these is a service that lets a holder swap one asset for another or shift value from one blockchain to a different one, and each leaves a trace on a public ledger that analysts can follow. The volume and speed of the activity raised an obvious question for investigators: whether software that can decide for itself where to send money was making the choices.
Elliptic's answer, so far, is no. The firm described the laundering as automated, but it said nothing in the observed transactions shows an AI agent making decisions. That distinction matters for how investigators think about the threat, and for how exchanges and analytics firms prepare for the next large theft.
The finding arrives in the middle of the largest exchange theft on record, so even a narrow technical conclusion carries weight. Exchanges, regulators and security vendors are all trying to learn how the operation was carried out, and the answer shapes which defenses are worth building. A laundering effort driven by scripts calls for one set of countermeasures, while one run by adaptive software would call for another.
2. How do scripts differ from AI agents?
A script is a fixed list of instructions. It performs the same steps in the same order, and it keeps doing so even when something goes wrong. An AI agent is meant to behave differently: it can pursue a goal, interpret the result of each step and change course without a person spelling out what to do next.
Elliptic offered a practical example of the difference. An agent that hit a failed transaction or a blocked wallet could work out the cause and switch to another exchange or bridge. Elliptic said it has not observed that kind of adaptive behavior in the Binance transactions.
Tom Robinson, Elliptic's co-founder and chief scientist, said the transactions look more like conventional software carrying out instructions set by people than like an AI system choosing its own route. Automation, in other words, is not the same thing as machine judgment, and the Binance case appears to fall on the automation side of that line.
The distinction is easy to blur in public discussion, because the word automation is often used loosely. A scheduled program that swaps tokens every few minutes is automated, yet it has no understanding of its surroundings. An agent, by contrast, would be expected to read the outcome of each action and decide what to do next. Elliptic's test is therefore a behavioral one: how does the software respond when a step does not go as planned?
3. Evidence in the transaction patterns
The strongest clues come from how the hackers handled things that did not work. Elliptic pointed to repeated transfers and failed transactions that resemble scripts executing preset instructions. A system able to recognize why a transaction failed might have tried an alternative path. Instead, the activity shows the same actions being sent again.
The hackers also reused the same transaction amounts. Before moving larger sums through a new route, they sent small test transfers to confirm the route worked. Probing a path with a minor amount is a long-standing habit of careful operators, and it is consistent with a human-written routine that includes a testing step.
Taken together, the repetition, the identical amounts and the cautious testing paint a picture of an organized, rehearsed process. That is a different picture from one in which a model improvises as conditions change.
Failed swaps are a particularly useful signal for this reason. On a decentralized exchange, a swap can fail for many ordinary reasons, and a human operator would normally investigate before trying again. A script has no such reflex. It simply repeats, and that repetition across thousands of transfers is what Elliptic sees in the data.
4. Six bridges in a 100-minute window
One detail drew particular attention. On October 30, the day after the theft, six different cross-chain bridges were used for the first time within a single 100-minute window. Cross-chain bridges let funds travel between separate blockchains, which makes them a favored tool for anyone trying to put distance between stolen assets and the place they were taken.
Elliptic said that timing does not settle who chose those services. The evidence does not show whether human operators or AI systems picked the six bridges, so the firm drew no conclusion on that point. The finding about scripts rests on how the software behaved when transactions failed, not on which bridges were selected.
The caution is worth noting. Elliptic did not claim to have proven that no AI tool touched any part of the operation. Its position is narrower: the observable behavior does not require an AI explanation, and what it has seen fits human-supervised automation.
Bridges are attractive to launderers because they connect networks that otherwise do not talk to each other. Moving value across several of them makes the trail longer, since an investigator has to follow the funds through each bridge's contracts and then pick the path up again on the destination chain. Using six of them within 100 minutes suggests a prepared plan to spread the funds quickly, though Elliptic stresses that the plan's author is not identified by that fact alone.
5. Who is behind the theft?
The FBI attributed the October 29 theft to TraderTraitor, a hacking group tied to North Korea. The agency also asked crypto services to block transactions involving wallets linked to the stolen funds, a request intended to make it harder for the group to convert the proceeds into cash.
Elliptic's new findings concern the period after the theft rather than the break-in itself. They cover how the hackers handled the money that escaped the initial freezes and whether artificial intelligence played any part in that stage. Laundering is where stolen crypto either gets cashed out or gets caught, so it is the phase that investigators and exchanges watch most closely.
Understanding the method has practical value. If the laundering relies on predictable scripts, with repeated amounts and repeated failed attempts, those habits give analysts patterns to detect and wallets to flag sooner.
Attribution to a state-linked group also helps explain the scale of the effort. A team that has planned the theft and the cash-out in advance can prepare tooling, test routes and queue up its transfers before the first alert goes out. That preparation is consistent with the small test transfers and the reused amounts that Elliptic describes.
6. How much of the stolen money has been recovered?
Elliptic put the amount frozen or seized at about $2.85 billion, which is 44% of the stolen funds. That is a large share for a crypto theft, and it reflects the freezes carried out after the attack, including the freezing of stablecoins.
The remainder is still largely in the hackers' hands. Elliptic said they control about $3.28 billion spread across more than 950 wallets. Splitting funds across hundreds of addresses is a standard way to make tracking harder and to avoid concentrating assets where a single freeze could catch them.
Those two figures show the state of the chase. Nearly half the proceeds are out of reach for the thieves, while more than half remain in motion or in storage, still subject to monitoring and to the FBI's request that services refuse transactions linked to the wallets.
The recovery picture also shows why speed matters. Every hour that stolen funds remain in a single place gives exchanges and stablecoin issuers a chance to freeze them. The hackers' use of about 20 automated services in under three days was, in part, an attempt to outrun those freezes, and the 44% that has been frozen or seized indicates that the response kept pace for a substantial portion of the money.
7. Why the question about AI matters
The debate over AI involvement reflects a wider concern in the security community. If autonomous agents could manage laundering on their own, they could in principle react instantly to freezes and blocklists, trying new exchanges and bridges faster than any compliance team could respond. Defenders would then be racing software rather than people.
Elliptic's reading suggests that, for now, the Binance laundering does not work that way. The hackers' tools repeat failed steps, which is the opposite of the adaptive behavior an agent would be expected to show. That leaves room for defenders to anticipate and interrupt the process, because predictable software is easier to profile than a system that changes tactics.
The finding is also a reminder that automation by itself is nothing new in crypto theft. Moving funds through roughly 20 services in under three days is possible with ordinary scripting, provided operators have prepared in advance. The speed on display is evidence of preparation, not necessarily of artificial intelligence.
For exchanges and analytics firms, the practical lesson is to keep monitoring for both possibilities. Detection systems tuned to repeated amounts, repeated failed transactions and test transfers can catch script-driven laundering today. If adaptive tools ever do appear, those same systems would need to look for different signals, such as routes that change in response to a freeze.
8. What to watch next
Attention now turns to the roughly $3.28 billion that remains with the hackers across more than 950 wallets. Further freezes, seizures or movements of that balance will change the 44% recovery figure Elliptic has reported, and analysts will be watching whether the laundering routines keep showing the same repetitive, script-like habits.
Elliptic has not said it will revise its conclusion, but its assessment rests on the behavior seen so far. If later transactions show the hackers adjusting their route after a blocked wallet or a failed swap without human prompting, that would be the kind of sign the firm says is absent today. Until then, the working view is that people are directing the operation and software is carrying it out.