Binance Hack 2026: How $6.3 Billion Was Stolen and What Every Engineering Team Should Fix
On 29 October 2026, attackers moved about $6.34 billion in crypto assets out of Binance wallets in under three hours. It is the largest cryptocurrency theft on record, roughly four times the Bybit theft of February 2025. Binance published its initial findings on 3 November, and the FBI, together with the US Secret Service and IRS Criminal Investigation, attributed the theft to North Korea's TraderTraitor activity. North Korea denies involvement.
The detail that matters for every engineering team is how ordinary the entry point was. No blockchain was broken and no zero-day has been reported. The attackers reached the wallet signing service through developer laptops, a fake job interview, an internal AI tool and a deployment token. Most companies have all four.
The Binance hack in numbers
- Assets affected: about $6.34 billion, Binance's revised estimate.
- Frozen or intercepted: about $2.85 billion, with the help of stablecoin and token issuers. Reuters reported freezes of roughly $1.57 billion by Tether, more than $800 million by Circle and about $200 million by Ripple.
- Customer impact: Binance says it absorbs the loss. By 31 October it had replenished 98% of the affected token assets from its own reserves. Replenished is not recovered: the stolen funds are still with the attackers.
- Downtime: withdrawals were suspended for about 48 hours and reopened step by step from 1 November.
- Laundering: analysts tracked the funds across more than 20,000 wallets in at least 65,000 transactions, mostly through cross-chain swap services.
How Binance was hacked: the attack chain
The sequence below follows Binance's initial findings as reported by Fortune and analysed by StepSecurity. Binance has announced an independent review, so details may still change.
- June: a tampered installer. A wallet DevOps engineer looked for an AI coding tool through a search engine and installed a tampered Homebrew package from an ad. Binance detected it, replaced the device and rotated credentials. By then the attackers knew who held which role in the wallet team.
- September: a fake job interview. A second engineer was approached on a recruitment platform. After interviews, including a live video call, the engineer ran a take-home coding assignment on a work computer. Code that executed during setup replaced a legitimate development tool component and connected to an attacker-controlled server. The endpoint protection in place did not flag it.
- 20 October: lateral movement through an internal AI tool. From that laptop the attackers used an internal Binance AI tool to take control of the computers of two more engineers. Binance has not disclosed how the tool was used.
- 20 to 29 October: a deployment token. On the DevOps engineer's machine the attackers found a container deployment token.
- 29 October, 17:15 to 19:56 UTC: the theft. With that token they deployed a container, reached the internal wallet signing service and submitted 50 signing transactions outside the normal withdrawal process, using nonce manipulation and malicious EIP-7702 delegations. Transfers ran across Ethereum, Bitcoin, XRP Ledger, Solana and Optimism.
- 30 October, 00:10 UTC: containment. Binance cleared the malicious delegation on-chain. Suspending withdrawals alone had not been enough, because delegations already set on-chain stayed valid until revoked.
Was AI the cause?
Early speculation blamed AI. Binance's findings point the other way: the initial access was conventional social engineering, and the report does not name prompt injection, a model flaw or a commercial coding assistant as the cause.
The AI tool still mattered, in a way that is easy to miss. It was an internal system that could act on other people's machines. Once one laptop was compromised, the tool became the path to the next two. That is a permissions problem, and it applies to any internal automation with broad reach: AI agents, remote management tools, CI runners.
We integrate AI into companies as an AI Forwarder, and this is the part of the incident we would underline for anyone rolling out internal agents: an AI tool that can execute actions is a privileged system and needs the same review as a deployment pipeline.
Seven lessons for teams that do not run an exchange
- 1. A developer laptop is part of production. If a laptop holds credentials that can deploy, it is inside your security boundary. Review it like a server.
- 2. Treat every repository setup as code execution. Interview assignments, sample projects and unknown repos belong in a disposable environment without company credentials or SSO sessions. This attack pattern is not new: the 2024 DMM Bitcoin theft also started with a fake recruitment assignment.
- 3. Give internal AI tools an identity and a boundary. List which machines and identities each tool can reach. Require authorization per target. Log who started an action, under which identity, against which system.
- 4. Remove long-lived deployment tokens from laptops. Use short-lived workload identities issued by the pipeline, and keep approval for production deploys on a path that an interactive session cannot complete alone.
- 5. Put policy inside the signing or payment service. A service that moves value should enforce its own limits: amount, velocity, destination and multi-party approval, independent of the infrastructure that deploys it.
- 6. Verify that an incident is closed. The June compromise was detected and cleaned up, yet the knowledge gained from it fed the September approach. After any endpoint incident, ask what the attacker learned, not only what they touched.
- 7. Hunt for replaced tools. The malicious code hid behind a trusted development tool. Watch for changed binaries, new symlinks, unexpected child shells and outbound connections from developer tooling.
What to check this week
- Which credentials on developer machines can change production? Count them, then shorten their lifetime.
- Can any internal AI tool, bot or agent run commands on machines other than the requester's? If yes, who approved that scope?
- Is there a rule and a sandbox for running third-party code, including interview tasks?
- Does your endpoint detection alert on tampered package managers and developer tools?
- Would your team notice 50 unusual signing or payment requests within minutes? Who gets paged?
- Do employees know that recruiters with unusually attractive offers are a known attack channel?
Frequently asked questions
How much was stolen in the Binance hack?
About $6.34 billion according to Binance's revised estimate, which makes it the largest crypto theft on record. Around $2.85 billion was frozen or intercepted with the help of token issuers.
How did the hackers get into Binance?
Through employees' work computers. One engineer installed a tampered software package from a search ad, another ran a malicious coding assignment from a fake job interview. From there the attackers used an internal AI tool and a deployment token to reach the wallet signing service.
Are Binance customer funds safe?
Binance says it covers the loss from its own reserves and had replenished 98% of the affected token assets by 31 October. The stolen funds themselves have not been recovered, apart from the frozen portion.
Who is behind the Binance hack?
The FBI, the US Secret Service and IRS Criminal Investigation attributed the theft to North Korea's TraderTraitor activity, the same cluster the FBI linked to the Bybit theft in 2025. North Korea denies involvement.
Did AI cause the Binance hack?
Not according to the initial findings. The entry was social engineering. An internal AI tool was then misused to move from one compromised laptop to two others, which makes the tool's permissions the issue, not the model.
Bottom line
The largest crypto theft so far did not need a cryptographic breakthrough. It needed one engineer to run an assignment, one internal tool with too much reach and one token that lived too long on a laptop. Each of those is a routine finding in a security review, and each is cheap to fix before an incident.
Want this checked in your environment? Our Pentest and SecOps work covers developer endpoints, pipelines and deployment credentials, SOC as a Service watches for the signals described above around the clock, and we design internal AI integrations with scoped identities from the start. Talk to us.