What the $6.34B Binance Hack Teaches Companies About Internal Security
How software companies can audit third-party tools, tighten onboarding, monitor privileged access in real time and secure internal AI platforms.
Written byMuhammad Abdullah Chaudhary
Key takeaways
Audit third-party tools
Every installer, package and AI coding tool on a developer's machine is part of your supply chain. Know what is there, how it updates and what it can read.
Tighten onboarding
Give new developers a safe way to install tools, verify who you are hiring, and start everyone with the least access they need.
Watch privileged access live
An alert on the outcome comes too late. Alert on unusual use of tokens and deployment rights, and hold the riskiest actions automatically.
Treat AI and internal platforms as entry points
Anything that can reach code, machines or other people's sessions needs its own identity, tight permissions and a log it cannot edit.
What happened
The first interview was a live video call. The HR contact seemed real enough, even sharing a social media account full of personal photos and videos. The second round was with the CEO. Near the end came a routine request: download this assignment.
The company didn't exist. The engineer, who worked at Binance, downloaded the assignment to a work computer. On 29 October 2026, about $6.34 billion left Binance's hot wallets. Several tracing firms call it the largest crypto theft on record. Binance says about $2.85 billion has since been frozen or intercepted, and that users will not bear the loss.
The interview wasn't the attackers' first move. By Binance's account, an engineer searching for an AI coding tool installed a tampered package in June, which the company caught. After the interview in September, the attackers used an internal AI tool to reach two more computers, found a deployment token, and got to the wallet signing service.
That account comes from Binance's initial report, and an independent review is still to come.
Who is behind it
The FBI attributes the attack to TraderTraitor, a North Korean group also tracked as Lazarus Group, and TRM Labs and Chainalysis reached the same view. North Korea denies any involvement.
The FBI has named the same group before: for the $1.5 billion Bybit hack in 2025, and for a 2024 scheme in which a fake LinkedIn recruiter sent a pre-employment test to an employee at the Japanese software company Ginco. That test led to the $308 million theft from DMM Bitcoin.
Why this matters if you don't run an exchange
Across the 33 engagements in our portfolio, we have built internal platforms, AI tools and data systems for clients from healthcare to finance to logistics. None of them is a crypto exchange. We are writing about this hack because the route looks familiar, and the defences are ones we build every day.
Until the very last step, nothing about this attack depended on crypto. It relied on developer laptops, a package manager, a job interview, a deployment token and an internal tool. Most software companies have all of those, along with their own version of a signing service: the production deploy, the payments API, the admin path into the customer database.
Here is the route. The numbered steps are Binance's account. The shields are where we would put a control.
The attack path
From a laptop to the signing service
6 chances to break the chain
June
Tampered install
An engineer searching for an AI coding tool installs a tampered Homebrew package. The device is replaced, but the wallet team is now exposed.
Chance to break the chain: Warn and watch the people exposed
September
Fake job interview
A fake HR contact and a fake CEO build trust with a second engineer, then send an assignment.
Chance to break the chain: Keep outside code off work machines
September
Malware on a work computer
The assignment runs on the engineer's work computer, and endpoint tools miss it.
Chance to break the chain: Limit AI tools to each user's access
20 October
Two more computers taken
Using an internal AI tool, the attacker takes over two more engineers' computers and finds a deployment token.
Chance to break the chain: Use short-lived tokens bound to one device
29 October
Container deployed
Stolen credentials deploy a container, which triggers a malicious script.
Chance to break the chain: Require a second approval for sensitive deploys
29 October
Signing service reached
The attacker can now start transactions outside the normal withdrawal process.
Chance to break the chain: Make the service enforce its own limits
29 October
$6.34 billion moved
Five hot wallets send about 233 million XRP in four minutes before Binance's internal alert, and more assets follow.
Those shields come down to four areas of work, covered below. But start with one small test.
One thing to do this week
Pick one deployment token from a staging environment, with the approval of whoever owns security. Copy it to a second machine you control, try a harmless read-only command, then revoke the token.
If the command works, you have the same kind of weakness the attackers used at Binance: a credential that works wherever it ends up. Section 3 covers the fix. If it is refused, good. Repeat the test on the next most sensitive credential you can safely check.
1. Audit your third-party tools
Binance's first compromise began with a search engine, and its report points to a SANS write-up of a fake Homebrew page spread through search ads. When the approved way to get a tool is slow or unclear, engineers find their own way, and that is how a tampered package ends up on a machine with production access.
Start with an inventory of what is on developer machines, including AI coding tools and their extensions. Then ask three questions of each one:
- Who publishes it? A named vendor you could contact, or an unknown account?
- How does it update? Automatic updates mean someone else can change what runs on your machines.
- What can it read? Source code, credentials and other people's sessions all raise the stakes.
An audit is a snapshot, so back it with a rule: tools come from sources you have checked, with pinned versions and checksums.
Every tool on a developer's machine is part of your supply chain.
2. Validate your onboarding process
Onboarding sets the habits and access people keep for years, so it is the cheapest place to get this right. We would check four things:
- Tools. A setup script that installs the toolchain from approved sources removes the reason to search for it.
- Identity. Chainalysis describes North Korean operatives posing as job applicants, using VPNs and stolen identity documents. This is separate from the Binance attack, but the defence belongs here: verify documents, confirm work history directly, and treat requests for unusual payment arrangements as a warning sign.
- Access. Start everyone with the least they need. Make sensitive access something people request, and have it expire.
- Awareness. Chainalysis says fake interviews target employees with privileged access. Tell those people, keep outside code off work machines, and offer a no-blame way to report a strange recruiter, because people hide an interview they think could get them in trouble.
The Binance engineer was an existing employee, so a first-week talk would not have been enough. Repeat the message whenever someone gains sensitive access.
3. Monitor privileged access in real time
Binance's alert fired on large transfers, after five wallets had already sent their XRP. By then it was too late to prevent them. The better signals came earlier: a deployment token sat on a laptop for nine days before it was used, and when it was used, it was from a different engineer's computer.
Write down what normal looks like. This is the step teams skip, and you cannot flag what nobody has defined. It is also where our best monitoring work starts. For a US revenue-cycle operator exchanging hundreds of secure files a day, we wrote down what should arrive, from whom and when, then built a watcher against it. Missing files that once surfaced as client complaints are now caught before clients notice. Privileged access works the same way: record who can deploy where, and from which machine.
Alert on access, not just outcomes. Good signals include:
- A token used from a device it was not issued to.
- A person deploying to a sensitive environment for the first time.
- A new container calling a service it has never used.
Hold, don't just notify. Most of the money left in about 20 minutes, according to TRM Labs. At that speed, the riskiest actions should wait for a person to release them.
Make the sensitive service enforce the rules. A process only protects you if the system behind it refuses everything else. Put permission checks in one central place that every protected path goes through, so no feature can skip them. That is what we built for Tasukaru: a single-company healthcare platform became multi-tenant SaaS, with tenant isolation enforced on every protected request, without rewriting the product from scratch. Pair that with short-lived, narrowly scoped deployment credentials issued through your pipeline. We build this kind of setup in our DevOps and CI/CD service.
Rehearse the stop button. Suspending withdrawals could not touch authorizations that lived on-chain, and clearing them took about four hours. List everything that grants standing access outside your application, such as OAuth grants, API keys, webhooks and scheduled jobs. Then run a drill that revokes them, and time it.
4. Treat AI and internal platforms as entry points
AI appears twice in this story: the coding tool that was the lure in June, and the internal AI tool used to reach more computers in October. Binance does not explain how that tool was used, and it describes no flaw in a model, so this is not proof that AI was the problem. It does show that an internal tool sat on the attackers' route, and Binance has since rebuilt that tool and its access controls.
Internal tools tend to get less scrutiny than customer-facing ones, while often having more reach. That applies to CI/CD systems, admin panels and deployment platforms as much as AI. Here is what we would put in place, based on systems we have shipped:
- Scope permissions tightly. Make assistants read-only wherever you can. Our on-premise data assistant runs a 30-billion-parameter model on the client's own server and answers questions in seconds, while staying read-only, permission-scoped and 100% on-premise, with zero data egress. If an assistant is misused, its permissions set the limit on the damage. OWASP lists this risk as excessive agency.
- Keep a person in the decision. In the AI systems we deliver, from care scheduling to disability support records, the model drafts and a named person decides. Actions on someone else's device or account should never run on the model's say-so.
- Give it its own identity. It should act with the requesting user's permissions and no more.
- Log every action somewhere the tool itself cannot edit.
Permission-scoped AI access is part of our AI security work.
Five questions for your next security review
| Question | The answer to worry about |
|---|---|
| Would a token copied from one laptop work on another? | Yes |
| How does a new developer install their tools? | "Search for it and follow the README" |
| What can our internal AI tools reach on a user's behalf? | Nobody is sure |
| Can anything reach production data or payments outside the normal flow? | "Only in emergencies, and we don't log those" |
| How long does it take to revoke every credential for one engineer? | We have never timed it |
Where to start
You do not need a large security team for this. Most of it is access design, and it is far cheaper before an incident than after one. It is also work we know well: we have delivered central access control, proactive monitoring and private AI in production for clients in healthcare and finance.
If those five questions gave you uncomfortable answers, a scoped technical audit is a sensible first step, and our team would be glad to talk it through.
Sources and further reading
- 29 October 2026 Binance Security Incident: Initial Findings and Response Update, Binance
- Binance Says Hackers Used Well-Known Tricks to Carry Out a $6.34 Billion Heist, Wired
- DPRK Hackers Perfected Their Fake Interview Playbook Over Dozens of Attacks Before Binance, Chainalysis
- The $6.4 Billion Binance Hack: Tracing the North Korea Link, Chainalysis
- Malicious Ad for Homebrew Leads to MacSync Stealer, SANS Internet Storm Center
- LLM06:2025 Excessive Agency, OWASP
- North Korea Steals USD 6.36 Billion From Binance in the Largest Hack in History, TRM Labs
- FBI links $6.3 billion Binance crypto heist to North Korean hackers, BleepingComputer
- What to know about the $6.3B hack on Binance, PBS NewsHour
- North Korea denies links to Binance hack as researchers trace stolen billions, CNBC
Related case study
Tasukaru Healthcare: Multi-Tenant Auth from Single-Company Software
