The theft of over $6.3 billion suffered by Binance on October 29 was reportedly not the result of a single breached computer or an isolated error. According to the reconstruction released by the exchange, the operation unfolded over months through a targeted campaign against technical staff: a tampered installer, outreach disguised as professional opportunities, and a malicious coding exercise allowed the attackers to gather intelligence, maintain access, and ultimately reach the systems that authorize wallet transactions.
It is a significant case well beyond the crypto sector, both for the financial scale of the attack and for the methodology. The incident illustrates a compromise that combined tools typically associated with social engineering campaigns with progressive reconnaissance of the target organization. In the decisive phase, according to Binance, the intruders also leveraged their own internal AI-based tool to expand control over accounts and access the wallet infrastructure.
A multi-phase campaign
The first milestone in the timeline provided by Binance dates back to June. A DevOps engineer working on wallets reportedly searched online for a well-known AI-assisted coding tool and downloaded a compromised Homebrew installer, apparently delivered via a fraudulent ad in search results. That software exposed details to the attackers regarding the target's role and colleagues connected to the wallet infrastructure.
The company states that it detected suspicious activity at this stage, replacing the engineer's device and rotating credentials as a precautionary measure. However, removing a compromised machine did not eliminate the intelligence advantage already gained by the orchestrators of the attack. Knowing which individuals hold technical responsibilities, how they collaborate, and which tools they use is often enough to make a subsequent impersonation campaign far more credible.
The second stage, in September, reportedly involved another engineer. The attackers posed as recruiters and crafted a hiring process spanning multiple calls and interviews, alongside social media profiles used to lend credibility to the offer. At the end of the process, the victim was asked to download a technical test: disguised as a coding exercise, it concealed malware.
Binance claims that the malicious software replaced the pyenv executable, a component used in Python development environments, with an altered version. This enabled it to execute commands and communicate with an attacker-controlled server. The company reports that the malware, described as a zero-day, featured obfuscation and evasion techniques that prevented it from being detected by endpoint defenses at the time. Steps were also reportedly taken to wipe traces of the intrusion and hinder the subsequent investigation.
From the endpoint to wallets, via internal AI
The most sensitive part of the account concerns the leap from employee devices to the ability to control funds. Binance has not released full technical details about its internal AI tool or the authorization architecture involved, but claims that the attackers used it to gain control of wallets associated with the first compromised engineer and a third employee.
By October 20, according to the company, the intruders had reached that level of access. On October 29, they reportedly logged into the third employee’s computer and launched the operation against the wallet infrastructure. Richard Teng, co-CEO of Binance, dismissed the idea that the theft hinged on a single laptop: his account describes a months-long progression in which access and information were chained together all the way to the system that authorizes transactions.
The distinction is not merely defensive. An attack traceable to a single stolen credential would require looking for a specific single point of failure; an operation carried out across multiple people, devices, and application layers, however, highlights an issue with privilege separation, behavioral monitoring, and the validation of sensitive actions. In a financial infrastructure, endpoint protection remains necessary, but it cannot be the final barrier before authorizing transfers of such high value.
Insufficient detail has been publicly disclosed to determine how the AI tool contributed to the escalation: whether through accessible data, automations, operational capabilities, or a combination of these factors. This limitation matters, as it prevents drawing generalized conclusions about enterprise AI applications. However, the case suggests that these systems must be included in threat models just like any other service with access to identities, documentation, workflows, or permissions.
Attribution and recovered funds
The FBI linked the attack to TraderTraitor, a North Korean hacking group. Teng acknowledged the connection cited by US authorities, without personally attributing the operation to North Korea. TraderTraitor is the moniker used for activities that have previously relied on job-related lures and fake career opportunities to target individuals working at crypto and tech companies.
Meanwhile, Binance pledged to cover customer losses using its own reserves. As of November 3, Teng stated that 98% of the $6.34 billion in affected assets had been restored, with less liquid assets still to be purchased. The exchange also reported that industry partners, including Tether, Circle, and Ripple, helped freeze approximately $3 billion in assets, while Coinbase reportedly blocked transfers to compromised wallets and provided direct assistance.
Freezing funds, however, does not equate to full recovery. A portion of the funds continues to move through laundering channels, and the effectiveness of the response also hinges on how swiftly exchanges, stablecoin issuers, and asset custodians can recognize and block the affected addresses. It is a mechanism that may work for certain assets and in certain jurisdictions, but it does not eliminate the risk of transfers toward services or assets that are harder to track and freeze.
Announced measures and open questions
Binance states that it has rebuilt the affected wallet system and its internal AI tool, replaced compromised devices, revoked credentials and tokens associated with the affected engineers, and expanded security scans to employee devices. It also temporarily restricted certain withdrawals. The company has announced an independent review of the measures taken—an important step, as the post-mortem published by the firm outlines its own account of the incident rather than a completed external audit.
The broader question of trust in major centralized intermediaries remains unresolved. The breach triggered a negative but brief reaction across crypto markets, according to CNBC, without the prolonged fallout seen in other platform collapses across the sector. Long-term resilience will depend on Binance's ability to complete reimbursements, clarify the conditions that enabled the escalation from employee machines to authorization systems, and demonstrate that its new barriers go beyond simple endpoint remediation.
For companies embedding AI-powered internal tools into operational processes, the takeaway is clear: these environments cannot be evaluated solely on productivity and response quality. Once they enter the chain managing identities, code, confidential data, or critical workflows, they become part of the attack surface. Security must therefore enforce least privilege, independent checks on high-impact actions, and monitoring capable of spotting anomalies even when an attacker leverages legitimate tools and seemingly valid credentials.
