Every Binance breach, from $41M to $6.3B: eight years, five incidents, and these 96 hours

TL;DR

• On 2026-10-29 at 21:18 UTC, CZ confirmed Binance had been robbed of about $6.3B, the largest single theft in crypto history and about 154x the 2019 incident ($41M).
• As of 2026-11-02 05:30 UTC: withdrawals resumed in stages on 11-01 (15 networks, $BTC first), 90–95% of the stolen assets have been replaced from the firm's own balance sheet, and about $3B is frozen.
• The problem was not a leaked private key but a compromised internal signing and approval chain, then EIP-7702 delegation used to sweep the hot wallets in bulk. The FBI, TRM Labs and Chainalysis all point to North Korea's TraderTraitor (Lazarus).
• User losses on paper are zero; the cost moved onto the exchange balance sheet and its credibility. Net outflow in the 24 hours after withdrawals reopened was $3.64B–6.33B depending on the method, roughly 2–4% of its $171B in customer assets.
• The thing to watch is not whether the coins come back, but whether the 11-02 proof of reserves and the 11-03 ZK verification turn "fully backed 1:1" from a verbal promise into a verifiable fact.

Data as of 2026-11-02 05:30 UTC. Prices are Binance spot $BTC/USDT and $BNB/USDT. The timeline is built from the publication timestamps of public reporting and official announcements; the exact second-level ordering of on-chain actions follows the transactions themselves. Two percentages in the source material disagree with the prices it gives and were recomputed; see section 6.

1. The long view: eight years, five incidents, a widening attack surface

Binance's security history has one clear through-line: every incident happened at the hot edge — hot wallets, APIs, a cross-chain bridge, the signing service. The cold wallets were never breached head-on. The scale went from tens of millions to billions.

Two things deserve to be called out separately.

First, those 7,000 $BTC from 2019 are worth about $548M at today's $78,336 — the same coins, up 13x in seven years. That is why $41M sounds small in its own era and belongs to a different order of magnitude on today's balance sheet.

Second, 2022 was strictly a BNB Chain ecosystem bridge exploit rather than exchange accounts being drained, though the market habitually files it under Binance.

Phished APIs in 2018, stolen credentials in 2019, a contract verification flaw in 2022, compromised signing infrastructure in 2026 — every time, the attacker went around the cold wallets and came in at the layer closest to withdrawal.

2. 2026-10-29: the 96-hour timeline

It took 5 minutes from the first on-chain anomaly to official confirmation, about 10 hours from confirmation to attribution as a state-level attack, and nearly 60 hours from the withdrawal freeze to withdrawals actually reopening. All times UTC.

3. Method: how one laptop opened nine hot wallets

Three pieces on this site already take this layer apart, so this is the skeleton; the detail is in those.

• The entry point was a person's device, not a key. Binance accepts the attacker may have entered its internal infrastructure through a compromised DevOps laptop. The BBC quotes a Binance co-CEO describing how one laptop opens one internal system, then another, until it reaches a service trusted to issue withdrawals — withdrawals that travel the legitimate path and look routine BBC. Blockaid put it more bluntly: the signing infrastructure was compromised, and they do not believe keys leaked. The full root-cause breakdown is in Binance names the root cause: what broke was a signing-capable service identity, not EIP-7702.
• Execution was amplified by EIP-7702. The attacker packed 50 EIP-7702 authorizations into one transaction; one of them had a nonce matching the target account and was accepted, delegating Binance 14 to an unverified sweeper contract. Bonanza Factory estimates the abused internal signing authority touched 9 hot wallets, and TRM traces funds from 10 chains. The preparation and the on-chain evidence are in The EIP-7702 attack chain: 18 hours of preparation, one signature, one bot, and how the signed payload becomes account code is in The EIP-7702 signing flow: from a hash to rewritten account code.
• The first cut was $XRP. TRM's reconstruction shows the attacker emptied 233.46M $XRP (about $296M) first, in 4 minutes Bloomingbit. That was not arbitrary: $XRP is deep and has many cross-chain routes, which makes it the best thing to move before anything can be frozen.

This is why replacing the coins matters more than recovering them. About $1.45B of the stolen assets is $BTC, which cannot be frozen at the protocol layer; the stablecoins are the interceptable part. So Binance covered the customer shortfall from its own assets while working with issuers to freeze what could be frozen — two legs at once.

4. Fund flows, replenishment and current status

The freezes landed fast, and that is the biggest operational difference from 2019: in 2019 almost nothing was recovered; in 2026 nearly half was frozen within 24 hours.

The outflow numbers differ by method and have to be read separately. The Block counts $3.64B in net outflow after withdrawals reopened. Cointelegraph cites $5.8B and stresses it is a small share of roughly $171B in customer assets. DefiLlama reads $6.33B over 24 hours ($5.86B at another cut). The spread comes from whether the window starts when withdrawals reopened or spans a fixed 24 hours, and whether internal transfers are netted out. They agree on direction: the outflow is large, and it did not become a bank run.

One more figure from CZ is worth noting: 98% of users chose to stay. That is a PR line, but it is the same order of magnitude as outflows being only 2–4% of customer assets.

One gap in the disclosure is still open. The missing items listed in Binance still has not published the signing logs — the signing-side logs and a post-mortem — have not been filled in as of this cutoff.

5. The market and credibility cost

The price reaction was violent but contained, and structurally unlike the panic of 2019.

• $BTC finished October −7.32%, its worst October in 12 years.
• $BNB, which trades as a proxy for exchange equity, fell more than $BTC — consistent with an event aimed at the exchange itself.
• Leverage flush: $1.53–1.6B liquidated across the market, a mid-sized washout rather than systemic deleveraging.
• Transmission into traditional finance: spot $BTC ETFs saw $643.2M of net outflow, and crypto-linked equities fell with it — Coinbase down over 13% (compounded by a Q3 revenue decline), Circle over 9%, Strategy about 8%, IREN over 10%.
• Narrative: Standard Chartered expects this to accelerate flow toward DEXes, while Bitwise's Matt Hougan says a theft this size would have sunk Binance and $BTC a decade ago and will not now.

Set against 2019: $41M then raised an existential question about whether the exchange would survive. $6.3B now raises an operational one about whether reserves are sufficient and whether withdrawals will clog. The market now treats Binance as a financial institution to be stress-tested, not a website that might disappear.

6. What this method cannot tell you

Two percentages in the source material disagree with the prices in the same material and were recomputed.

• The $BNB figure of 6.2% below the pre-incident high became 6.5%. Using the two prices the material itself gives, $710 ÷ $759 − 1 = −6.46%. For 6.2% to hold, the pre-incident high would have to be $757, not $759. Only one of the two can stay, and this piece keeps the price and recomputes the percentage.
• The $BTC claim of a $77,331 low on 10-31 and a −5.0% drawdown was not used. The same material's own timeline quotes $76,771 at 10-30 15:03, below that supposed low. So this piece records $76,771 as the lowest reading, which is −5.7% from $81,400, and timestamps every reading rather than asserting which was the true low of the window — the material never gives intraday lows.

Four more limits belong to the method itself:

• The timeline is reporting time, not chain time. The timestamps come from when announcements and articles were published; the actual on-chain ordering follows the transactions, and the two can differ by minutes to tens of minutes.
• The roughly $3B frozen does not add up from its parts. Tether $1.567B plus Circle $800M+ plus Ripple about $200M is $2.567B, with the remainder filed under project-level freezes and no breakdown. The number is a stitch-up of separate disclosures, not a reconciliation.
• The total itself has four versions ($6.3B / $6.36B / $5.84B / $5.3B), differing by cutoff time and pricing method. This piece uses the Binance figure of $6.3B, but every ratio built on it carries that uncertainty.
• Proof of reserves has not been published, so "fully backed" is not verifiable today. All that can be said is that Binance claims 90–95% replenishment, not that it has been replenished — which is exactly why the first watch item exists.

Conclusion. Across eight years, the Binance security model has worked, but it defended the wrong thing. The cold wallets were never breached, SAFU never defaulted, and compensation was never partial — on the axis of coins being stolen, the system succeeded. What 2026 proves is that when the attacker switches targets from the vault to the pen that signs withdrawals, hot-cold separation, multisig and the insurance fund all stop mattering, because the transactions look entirely legitimate to the system.

Three things actually changed. First, the scale crossed the threshold of what a company absorbs and a market ignores: $6.3B is 3.7% of customer assets, covered only by stacking SAFU, the futures insurance excess and the treasury. That is not a reusable buffer. Second, the adversary is now a state-level APT — Lazarus has taken $11.87B cumulatively — so the thing being defended against is patient, funded and capable of supply-chain intrusion. Third, trust is priced differently: the market no longer asks whether you will pay, it asks what proof you have that you can. That is why the 11-02 proof of reserves matters more than any statement.

Bottom line. This will not end Binance, but it ends the old frame of an exchange hack as a black swan. From here, an exchange is anchored on verifiable reserves and withdrawal resilience, not on volume.

Three things to watch over the next 72 hours: first, whether the full 11-02 PoR covers every stolen asset and includes the $6.34B provision — majors only counts as incomplete; second, whether the 11-03 ZK verification lets users check for themselves, which is the line between "trust me" and "verify me"; third, whether withdrawals slip again or pick up new limits — one more delay turns that $3.64B outflow into the start of a second run.

The first record of this event is Binance hot wallet compromise: one legitimate signature, 8 minutes, $5.64B. Research notes, not investment advice. DYOR.