Binance disclosed a multi-stage compromise and estimated $6.34 billion in affected assets.
The public account names tampered developer tooling, a fake coding assignment, compromised engineer systems and EIP-7702 wallet abuse.
What happened?
Binance’s November 3 incident account describes a multi-stage compromise that culminated in the October 29 theft of roughly $6.34 billion in digital assets. The exchange says about $2.85 billion had been frozen or intercepted and that it had replenished 98% of affected tokens as of October 31. Those are Binance’s figures, subject to reconciliation and market prices; independent on-chain analyses place the theft in the same broad range while differing on exact totals and the amount frozen.
The preparation began months earlier. Binance says a Wallet DevOps engineer downloaded a tampered Homebrew installation while looking for a popular AI coding tool. The company detected that event, replaced the device and rotated credentials, but says the attackers learned about the engineer and wallet team. In September, a second engineer was drawn into a fake recruitment process and asked to download a coding assignment. The repository executed during setup, replaced a legitimate developer-tool component with a rogue one and opened an attacker channel.
SecurityWeek reports that on October 20 the attackers used a compromised computer and an internal Binance AI tool to gain control of the first and a third wallet engineer’s systems. The report explicitly says Binance did not explain how the AI tool was used. Later, stolen credentials and container-deployment access reached a wallet-signing service. On Ethereum, attackers abused EIP-7702 authorizations so a malicious contract could obtain partial execution rights over a hot wallet, allowing transfers to continue beyond ordinary withdrawal controls.
Source: Binance ↗ · Source: Binance ↗ · Source: SecurityWeek ↗
The evidence supports a supply-chain and identity story, not AI-assisted coding caused it
The phrase “AI-assisted coding breach” collapses different events. One AI coding tool was the context for a tampered download in June. A separate internal AI tool appears in Binance’s account of the October 20 step, but its action is undisclosed. The fake coding assignment was a developer-workflow lure that installed or replaced tooling. None of those facts demonstrates that an AI assistant generated malicious code, autonomously moved funds or made the causal decision to attack the wallet infrastructure.
The strongest supported causal chain is narrower: attackers used social engineering and compromised developer environments to obtain access and knowledge; a rogue component opened a channel; credentials and deployment permissions reached a signing service; and on-chain delegation allowed asset movement to continue after Binance paused normal withdrawals. SecurityWeek and Binance align on the staged nature of the incident. Chainalysis and Elliptic add independent tracing, while the FBI separately attributes the activity to North Korea-linked TraderTraitor. Attribution is separate from technical cause and is reported here as attribution, not a judicial finding.
EIP-7702 matters because the incident was not simply a stolen-private-key story. Binance says its multi-party-computation wallet had no single key that could immediately revoke the malicious authorizations. Independent analyses describe a race between the attacker’s transactions and stablecoin issuers’ ability to freeze assets. The lesson is not that EIP-7702 is inherently a breach mechanism; it is that new authorization paths can create on-chain state that survives controls designed for an exchange’s normal withdrawal system.
Source: Binance ↗ · Source: Binance ↗ · Source: SecurityWeek ↗ · Source: Elliptic ↗ · Source: Chainalysis ↗ · Source: FBI ↗
Why does it matter?
For engineering teams, the practical warning is about trust boundaries around developer tools. A popular installer, a repository used in a hiring exercise, a user-wide agent instruction file or an internal automation tool can all become part of the attack surface when they inherit credentials and context. Endpoint detection may miss a component that looks like a normal dependency, while resetting one laptop may not undo information learned before detection. Organizations should isolate coding exercises, use least-privilege credentials and require independent approval for deployment and signing operations.
For AI security reporting, the disclosure is a test of causal discipline. The internal AI tool is newsworthy because it appears in the attack chain, but the public account does not reveal whether it summarized, searched, executed, routed or merely exposed information. Filling that gap with a familiar narrative about an agent writing malware would be speculation. The same restraint applies to North Korea attribution: the FBI, Elliptic and Chainalysis provide relevant evidence, while Binance’s technical account tells us what the exchange says happened inside its systems.
Binance says it is rebuilding affected wallet systems, replacing devices and engaging independent security firms. That is a response plan, not proof that the root cause is closed. The durable story is a compound failure involving people, software supply chains, endpoint trust, credentials and blockchain authorization. AI tooling may have been one component in that chain, but available evidence does not justify making it the cause.
Source: Binance ↗ · Source: Binance ↗ · Source: SecurityWeek ↗ · Source: Elliptic ↗ · Source: Chainalysis ↗ · Source: FBI ↗
Read the sources
- Binance incident public post ↗
Binance · November 3, 2026 - Initial findings report ↗
Binance · November 3, 2026 - Rigged coding assignment opened path to Binance heist ↗
SecurityWeek · November 4, 2026 - Largest theft in history: Binance hack ↗
Elliptic · November 3, 2026 - Tracing the $6.4 billion Binance theft ↗
Chainalysis · October 30, 2026 - FBI TraderTraitor alert ↗
FBI · October 31, 2026
