A Fake Job Interview Opened Binance's Door: $6.3 Billion Gone

What Happened

According to a report from Binance (reported by The Register), the attack chain was broken down into five steps:

1. June Alert. An engineer at Binance downloaded a malicious Homebrew installation package, which was part of a supply chain attack. The security team detected it, and all affected devices and credentials were rotated. However, according to Binance, information about Engineer A’s role and colleagues in the wallet infrastructure team had already been exposed to hackers.

2. September Impersonation Interview. Another engineer was approached on a legitimate job application platform by someone claiming to be a project director, who arranged a series of interviews. In the first video interview, the fake HR shared a social media account with personal photos and videos to build trust. Binance reported that this was more sophisticated and deceptive than any social engineering tactics they had seen before. The second interview was conducted on a popular social chat app by a fake CEO, who instructed the engineer to download a “task” containing malicious code from a repository. This task was executed on his work computer.

3. What the Malicious Script Did. During the setup process of the “task” repository, it replaced pyenv 2.8.0 in Homebrew Cellar with a malicious version: first writing pyenv_ (3,414 bytes), deleting the legitimate pyenv (3,344 bytes), renaming the malicious version to replace it, then removing /opt/homebrew/bin/pyenv,重建 symbolic links, and starting zsh. Binance stated that endpoint security tools did not immediately detect this malware, and the attackers also took measures to hide scripts and evade detection.

4. October 20th Lateral Movement. The attackers used this controlled computer along with Binance’s internal AI tool to gain access to the computers of DevOps engineer A and another wallet software engineer C.

5. Action on October 29th. The attackers logged into Engineer C’s computer, used Engineer A’s container deployment permissions to deploy tokens, and triggered scripts to steal funds.

The report identified five aspects of the attack: impersonation social engineering, malicious downloads, compromised internal AI tools, endpoint takeover, and a sophisticated transaction authorization method (specific implementation not disclosed). Binance concluded that it was a complex multi-stage operation, not a single-point intrusion or simple phishing.

June Alert Did Not Close the Door for October

Critics argue that the main issue was not the sophistication of the hackers, but rather a process audit to identify where the chain was broken.

First, successful detection does not mean completion of action. The malicious installation package was caught in June, devices were replaced, credentials were rotated, and actions were taken, but the most critical step was missing: the exposure surface was not cleared. The adversaries already knew who Engineer A was, what his role was, and who was in the team. What was detected was a file-level intrusion, and what was leaked were organizational relationships. The lateral movement four months later was carried out along this line. If judged by “malware cleared,” June was considered successful; if judged by “what the adversaries obtained was invalidated,” that disposal had not been completed at all.

Secondly, no one stops engineers from running unfamiliar code on work computers. Engineers used company equipment to perform personal job-related tasks, and the endpoint tools did not immediately detect and replace the scripts that modified pyenv. Comments divided into two camps: some were shocked at how engineers used work computers for personal work; others questioned why this was unusual, as many colleagues also did unrelated things at work. This behavior is widespread, so defenses cannot rely solely on individual awareness. The boundaries of BYOD and minimum permissions are old issues, and the cost of this incident is estimated at $6.3 billion.

Thirdly, internal AI tools have become a new attack surface. Attackers used Binance’s own internal AI tools to gain access to the computers of two wallet engineers. In corporate networks, AI tools are new members with extremely high permissions, and the paths of compromise may not have been anticipated in traditional threat models.

Fourthly, job platforms have outsourced identity verification costs to individual applicants. Fake recruiters operated on legitimate job platforms for weeks to months, starting in September and succeeding in October, until the incident became apparent. Multiple comments on HN indicated that reports from fake recruiters on LinkedIn went unaddressed (according to the commenters). If the platform does not verify those entering, then each user must verify themselves.

Fifthly, the dehumanization of the recruitment process has led to this scam. Real-person interviews became a rarity, and applicants became less cautious toward “real” people who approached them voluntarily. The scammers did not break through defenses; they entered through doors left open by the defenses.

Disparities in Accounts

The amounts are listed as follows: The FBI and The Register claim $6.3 billion; TRM Labs claim $6.36 billion, with most of the funds withdrawn within about 20 minutes; Elliptic claims nearly $6.43 billion, and has confirmed that $2.85 billion stolen assets have been frozen or seized. The figures from different agencies are inconsistent, and this article does not take a single precise value as definitive.

The FBI’s alert was reported in accordance with official statements: The alert published a list of wallet addresses involved in the incident, urging exchanges, on-chain analysis companies, and virtual asset service providers to intercept them; hackers have begun converting the stolen funds into Ethereum, and investigators expect the funds to be transferred through DeFi platforms and instant exchange services to make tracking more difficult. TRM Labs and Elliptic, two on-chain intelligence companies, made similar attribution statements before the FBI’s announcement.

According to The Register’s article on the FBI’s attribution, Binance has now resumed customer withdrawals after the attack. The exchange claimed to have recovered about 95% of the stolen assets, with the remaining portion being difficult to recover due to limited liquidity; co-CEO Richard Teng promised that customers would not bear any losses. The subtitle of the article stated that during the emergency funding replenishment, customers withdrew billions of dollars. All of this is according to the article’s account.

The post-incident actions reported in Binance’s incident report include: rebuilding the compromised wallet system and internal AI tools; replacing all affected devices; canceling all credentials and tokens of the affected engineers; cooperating with law enforcement, blockchain intelligence companies, and security researchers to track and freeze assets; hiring independent cybersecurity experts for evaluation; and announcing the establishment of a security research fund focused on AI security in the blockchain environment. The report describes what was done, but there are two key questions unanswered: why the June actions did not cover the organization’s exposed relationships and what the permission model of internal AI tools looked like.

What Are the Comments Saying?

There are over 300 discussions on HN, with high-starred comments summarized into five main points:

I. Structural reflection on the recruitment ecosystem tops the list. The most praised comment suggests that the real tragedy is that job seekers are so eager for a face-to-face interview that they abandon common sense; employers complain about not being able to recruit and encountering scammers, but this dilemma is caused by their own actions: continuously streamlining processes and greedily screening, turning human interaction into a scarce resource. Some sarcastically say that tech companies make candidates go through various automated stages before seeing them in person, and now they get the same result. Others recall that in the 1990s, during interviews in Manhattan, candidates were required to take a written test first, and they refused to proceed on the spot.

II. Resentment over job-seeking platforms. One person shared that after posting their job status on LinkedIn, they received contact from multiple fake companies offering fake interviews. After registering with one consulting firm, they also received spam ads. Another said that someone impersonated their company’s recruiters and conducted long-term phishing attempts. After hundreds of reports went unaddressed, they finally got help by meeting an LinkedIn employee at a party and using internal channels to have their situation marked. One person made a harsher conclusion: LinkedIn is a wasteland, and Microsoft’s leadership probably forgot they own it. All of the above are from the commenters.

III. Debate over boundaries of work computers. Some are confused about how engineers use work computers for private job-related tasks; others ask what’s so special about it; others offer technical solutions: such tasks should at least run in Docker containers, with network access and volume mounting controlled.

IV. Sarcasm and humor. Someone said that if the task was a security position test, the sarcasm would be maximum.

V. The shift in corporate security practices. Some observed that several companies are allocating security operations resources to review candidates, claiming it’s dangerous outside.

Three things ordinary users can do

1. Run interview tasks only in isolated environments. Do not trust any downloaded test questions, code repositories, or testing tools: execute them in virtual machines or containers, control network access and volume mounting, and destroy the environment after completion; never run them on computers with real data and credentials, especially work computers. The entry point for this is a single setup execution.

2. Reverse verify the recruiters, do not trust their provided identity materials. Fake HRs prepare social media accounts with photos and videos to build trust, but that doesn’t mean it’s real. Verification steps: use the official domain name instead of the given link; confirm the recruitment contact by replying to the official email or phone; search for the company name with “scam” keywords to see precedents; the more they urge you to download the task, the longer you should pause.

3. Separate private job applications from company equipment, and have management include AI tools in permission audits. Job seekers: Using private computers for interviews and tasks is a basic rule, not a matter of cleanliness. Management: Include “using company equipment to execute external code” in isolation and approval policies; assess the permissions of internal AI tools and monitor them by level. This horizontal move follows a path that no threat model has yet included.

Conclusion

Disclosure Statement: This article does not make unfounded assumptions or accusations. The full report on the Binance incident has not been independently verified, and details of the attack chain are quoted from The Register. The text is marked according to the statements provided by the disclosing parties; FBI attribution is a quote from official statements, without direct citation or interpretation. Amounts are taken from the ranges reported by various organizations; any identification of job-seeking platforms in the comments section is the perspective of the commenter. There are three things that can be confirmed: both timelines have their sources; the amounts reported by different organizations are indeed inconsistent; the incident report lists five attack vectors, which are considered by the exchange. This case raises two real questions for the industry: when recruitment processes turn real people into a scarce resource, scams can exploit this scarcity; when internal security procedures treat June’s alert as a single event rather than an ongoing exposure that lasted until October.

Next, we will focus on two key points: first, the whereabouts of the stolen funds. The FBI has released a list of wallet addresses involved in the case. Hackers began converting Ethereum, and Elliptic has verified that the frozen and seized amounts total 2.85 billion. Any significant progress in money laundering or tracking will be followed in this section. Second, on Binance’s side: conclusions from independent expert reviews, detailed information from security research funds—any announcement will trigger version 2. The 48-hour window closes around 14:09Z on November 6th. If any changes occur during this window, version 2 will be updated.

Reference Sources

• The Register’s detailed article (2026-11-03, Jessica Lyons; local evidence ev1105-register-text.txt. The five-step attack chain, pyenv replacement script, and five attack vectors are quotes from the Binance report in this article, with its own statements.
• The Register’s FBI attribution article (2026-11-02, Carly Page; local evidence ev1105-fbi-text.txt. FBI alerts are quotes from official statements; TRM Labs’ 6.36 billion, Elliptic’s 6.43 billion, and 2.85 billion frozen data are second-hand references from this article; Binance’s recovery of withdrawals, approximately 95% compensation, and joint CEOs’ commitments to this article’s perspective).
• Binance Incident Report (published on binance.com blog on November 3, 2026; the full text has not been independently verified: no naked data, proxy 401, Wayback archive unavailable; all details in this article are quotes from The Register, without any upgraded statements).
• Hacker News story 50365755 (listed on November 4, 2026, at 14:09Z; Algolia snapshot 545▲/327💬, numbers still rising, the text uses more than 500▲ of the quoted data; 60 comments from ev1105-hn-nkbinance.json, all quoted in quotation marks).

Author: Yongliang