Also on Spotify and Apple Podcasts: search “Exploit Brokers by Forgebound Research”
Picture this. A recruiter messages you. Good company, good pay, remote. Two interviews go well. Then comes the take-home: a small coding test. You’ve done a hundred of these. You download it, you run it.
On November 3rd, Binance published its initial findings on the October 29th incident behind a theft of roughly $6.3 billion, and one of the early stages it describes is exactly that: a fake job interview that ended in a programming test. By The Hacker News’s numbers, that makes it more than four times the size of the Bybit hack, the previous record. (The “record” framing is the press’s and Elliptic’s wording; neither Binance nor the FBI makes that claim.)
This post walks through what Binance, the FBI, the DOJ and analytics firms have said publicly, and tries to keep three things separate: what is claimed, what is context, and what is still unknown. A quick standing note: I’m a Senior Software Engineer with working experience in security and AI, but I’m not your engineer or your security person. Everything here is my own opinion and general information.
At a glance
Entry point (per Binance): a fake job interview ending in a malicious “assignment,” plus an earlier tampered Homebrew installer.
Attribution: the FBI (October 31) says North Korea is responsible and calls the activity TraderTraitor. North Korea has denied it.
Amount: about $6.3B. Binance says $6.34B; other sources differ (see below).
Motivation (context from public reporting, not Binance’s report): crypto theft is widely reported as a way a heavily sanctioned state brings in hard currency.
How fake-interview campaigns work

The attacker builds a fake company, a fake recruiter, and sometimes a fake “CEO” for the second round. They find an engineer on a job platform, reach out, and run what looks like a normal hiring process:
- An HR screen
- A technical round
- A take-home assignment, a coding test the candidate completes on their own machine
The interview is the delivery mechanism.
Think about what the assignment asks you to do:
- Clone a repo
- Install dependencies
- Run it
That’s the job. There’s no strange attachment and no “enable macros” prompt. The malicious action looks exactly like the normal one, and the trust is already built into the workflow.
And if you’re picking a target, who do you want? Not a random employee. You want the wallet engineers, the people who sit close to signing and approval flows, which at a crypto company puts them close to the keys. That’s my read.
What Binance says happened: two earlier stages
Binance’s report describes two separate earlier stages against two different engineers. Some outlets blur them together; Binance keeps them apart.
| Stage | What Binance says |
|---|---|
| June · Engineer A (wallet DevOps) | Searched for a popular AI coding tool and downloaded a tampered Homebrew installer. Security caught the anomaly, replaced the device, and rotated related credentials. |
| September · Engineer B (software engineer) | Contacted through a legitimate job platform by someone posing as a project director. Round one: a fake HR video call. Round two: a chat app and a request to download an “assignment,” a repository containing malware. |
A take-home assignment is code from someone who hasn’t been verified. In Binance’s account, one developer’s machine was the foothold for a multi-stage campaign. This was a malicious assignment, not a documented compromised upstream package. In my own environment, I treat take-home tests the way I’d treat any untrusted code: in a throwaway VM, not on my main machine.
~Cipherceval
The pyenv swap
This is the part that, as a developer, made me put my coffee down. It concerns Engineer B, the September stage. According to Binance’s file log, the malware replaced a legitimate development tool component with a malicious one, and that component was pyenv. It:
- Deleted the valid
pyenvbinary - Swapped in a replacement with the same name
- Deleted and recreated the link under Homebrew
- Started a shell
Separately, Binance says the malware set up communication with an attacker-controlled server.

What pyenv and shims are
pyenv manages multiple installed Python versions so projects that need different Pythons can coexist. It places small launcher files, shims, on your PATH. When a Python command goes through a shim, the shim invokes pyenv to select the interpreter.
Why that tool? Binance doesn’t say
My inference, labelled as such: a malicious replacement in that path could get repeated chances to run while looking like a familiar tool. Binance hasn’t said that is why it was chosen, or how often the replacement ran.
“Zero-day” and what it does and doesn’t mean
Binance calls the malware zero-day and says its endpoint protection didn’t detect it at the time. That doesn’t, by itself, show the attackers exploited an unpatched software vulnerability. As described, the swap was a file replaced in a place the developer already trusts.
Toolchain files such as
pyenv sit on the PATH where many developers rarely look. Trusted hashes or signatures are one way to assess them. What I personally do is spot-check the things on my PATH that I never think about.~Cipherceval
Sideways to the wallet signing service

Binance describes the next steps this way:
- From Engineer B’s machine, the attacker moved to other engineers, including wallet engineers. Binance says an internal AI tool was part of how (next section).
- On Engineer A’s machine, the attacker found a container deployment permission token.
- On October 29th, in a window of about 2 hours 40 minutes by Binance’s times (no timezone given), the attacker logged onto a third engineer’s computer, used those credentials to deploy a container, and triggered a malicious script.
- That reached the wallet signing service, the system that authorizes the movement of funds.
The 50 signing transactions
Binance says the attackers submitted 50 signing transactions to, in its words, “manipulate the account nonce and establish malicious delegation authorizations.” Two terms help here. A nonce is a counter used to order account activity on Ethereum. EIP-7702 is an Ethereum feature that lets an account authorize code from a contract to run in that account’s context, and Binance says the attackers abused it.
The hot wallet (a wallet connected to the internet) used multi-party computation (MPC), where signing involves multiple parties rather than one readily available private key. Binance says it needed a special procedure to revoke the authorizations. That is what Binance says happened here; it is not a claim that MPC wallets can’t revoke delegations.
Binance hasn’t said how long the deployment credential remained valid, so I’m not assuming it was long-lived. My takeaway is narrower: a stolen deployment credential can be a bridge from a compromised laptop to a much more sensitive system.
A deployment credential on a developer machine was part of the path to the signing service. Short-lived credentials can reduce the window of exposure, though they can still be abused while valid.
~Cipherceval
The internal AI tool angle

This is where headlines run ahead of facts, so here’s the line between what’s said and what isn’t.
| Status | Detail |
|---|---|
| Said by Binance | On October 20th, the attacker used Engineer B’s compromised computer and one of Binance’s internal AI tools as a “springboard” to take control of two more engineers’ machines: the wallet DevOps engineer from June and another wallet software engineer. |
| Not said | Which AI system it was, or how it was abused. If you see someone online naming the product, I’d treat that as speculation. |
| Unknown | Whether the tool could run commands, access files, or reach other machines directly, or played some other role. |
More generally, and as my opinion rather than a description of the undisclosed mechanism: when an internal tool can act with privileged access, its permissions and who can use it matter.
I think of it as a permissions problem, not “AI is dangerous,” which is too vague.
The questions worth asking:
- What is a tool allowed to do?
- With whose credentials?
- Who gets to ask it to?
As AI tools get wired into engineering workflows, I think vendors and internal platform teams should treat their permission scopes with the same suspicion they’d give a service account. For teams that have rolled out internal AI tooling, what that tooling can touch is worth looking into. That’s a general observation, not an assessment of anyone’s environment.
Internal AI tools can widen a trust boundary. My opinion is that their permissions deserve the same scrutiny as any service account.
~Cipherceval
Attribution: the FBI and TraderTraitor
On October 31st, the FBI said North Korea was responsible and calls this activity TraderTraitor. Some outlets say that’s the same unit as the Lazarus Group; the FBI’s alert uses only the TraderTraitor name, so that’s the name I use here.
Two things worth knowing:
- It’s the FBI’s attribution. The accurate phrasing is “attributed to,” not “was.” Binance’s written report cites the FBI rather than making the claim itself, and North Korea has denied it: a senior official quoted by Bloomberg called the claim “absurd rhetoric.”
- The FBI doesn’t publish its basis. The analytics firm Elliptic, before the FBI announcement, assessed TraderTraitor as the likely culprit, citing funds laundered alongside an earlier theft it had assessed as North Korea-linked, services shared with earlier TraderTraitor thefts, and similar methods. That’s context, not proof.
The money: frozen, laundered, covered

The FBI’s alert, in brief:
- 628 wallet addresses listed: 608 Ethereum, 19 Bitcoin, 1 Solana
- The actors have been converting the stolen assets primarily into Ether
- The FBI expects DeFi and instant swap services next
- It urges exchanges and node operators to block those addresses
In my view, laundering is the second half of every one of these heists: stealing is step one, spendable money is step two, which is why freezes matter.
| Figure | Who | Notes |
|---|---|---|
| ~$6.34B stolen | Binance | TRM ~$6.4B; The Hacker News $6.43B; FBI/DOJ “more than $6B.” No source I found explains the spread. |
| ~$2.85B frozen or intercepted | Binance | Subject to reconciliation. |
| ~$2.4B still held by attacker | TRM Labs (Nov 3) | Moving funds through self-minted tokens and chained bridges. |
| >$2.7B frozen; >$1.2B court-authorized for seizure or restraint | DOJ (Nov 12) | The $1.2B is within the $2.7B. Not forfeited; recovery ongoing. A separate deepfake video-call scheme in the same announcement is a different case. |
| 98% of impacted token assets replenished (Nov 3); coverage completed with company funds (Nov 12) | Binance | Withdrawals reopening in phases. |
Binance also says it rebuilt the wallet system and the internal AI tool, replaced affected devices, decommissioned related credentials and tokens, and ran enhanced security scans across all employee devices. Good. In my opinion, the part most worth hearing more about is the credential model, because rebuilding systems is the easy part to announce.
The DOJ’s $1.2B figure is a lower bound inside its $2.7B frozen total, so it doesn’t establish a discrepancy with Binance’s roughly $2.85B figure.
~Cipherceval
What isn’t settled
- The amount: $6.34B per Binance; $6.4B to $6.43B in other reports; “more than $6B” per the FBI and DOJ. I use “about $6.3 billion.”
- The timeline: some reports say the first engineer was breached in June; others describe a September fake interview in two rounds. Binance’s report separates the two stages, as above.
- The AI mechanism: which tool, and how it was abused, hasn’t been disclosed.
- Credential lifetime: not disclosed.
- Attribution basis: the FBI hasn’t published it.
Five things worth knowing
🔢 Key Takeaways
1. 📋 Worth Knowing: the interview can be the attack. A take-home assignment is code from a stranger. By Binance’s account, one developer’s machine was a foothold for a multi-stage campaign.
2. 🔍 Things to Consider: your toolchain is part of your attack surface. The pyenv swap targeted a component many developers don’t routinely inspect. Trusted hashes or signatures are one way to assess toolchain files.
3. ⚠️ Keep in Mind: deployment credentials can cross a trust boundary. Binance says a token from a developer machine was part of the path to the signing service. Short-lived credentials can reduce the exposure window but can still be abused while valid.
4. 📋 Worth Knowing: internal AI tools can expand the trust boundary. Their permissions deserve the same scrutiny as any service account.
5. 🔍 Things to Consider: the facts aren’t settled. The exact amount and the AI mechanism remain open and subject to further reconciliation. The primary source is the place to check.
Underneath all five: security is the person and the technology. This started with a person doing their job and a process that looked normal. If this was useful, send it to a developer you know, especially anyone who’s job hunting right now. Nothing is hack-proof, but making it as hard as possible for the attacker is the whole game. We’re all in this together. Update your stuff, and that includes the tools you’ve stopped thinking about.
— Cipherceval
Sources
Primary
- Binance, “29 October 2026 Binance Security Incident: Initial Findings and Response Update” (Nov 3): link
- Binance announcement, coverage of impacted assets completed (Nov 12): link
- FBI alert, North Korea responsible for the Binance attack (Oct 31): link
- U.S. Department of Justice press release (Nov 12): link
- pyenv, “Understanding shims”: link
- EIP-7702: link
Analysis and reporting
- TRM Labs, Binance hack update (Nov 3): link
- Elliptic: link
- The Hacker News, FBI links TraderTraitor (Nov 1): link
- The Hacker News, how fake job interviews led to the hack (Nov 3): link
- CNBC, Binance co-CEO on the fake job interviews (Nov 3): link
- Bloomberg (Nov 2), North Korea’s denial: paywalled, no link.
Attribution is the FBI’s; Binance’s report cites the FBI. North Korea has denied it. Figures were current as of recording and may change.

Leave a Reply