Analisi on-chainPubblicato il 5 novembre 2026

The keys stayed safe. The money didn't.

What Binance, Bybit and Bitget teach financial institutions about access, people and signing.

Crypfy cover: Binance $6.3B, Bybit $1.46B and Bitget $388M lost in three exchange hacks, with zero private keys stolen.

In short

Binance, Bybit and Bitget lost about $8 billion without losing a single private key.

In all three, someone reached the system that asks for signatures and found it too open.

Below: what happened, what the cases have in common, what to check at your counterparties.

On 29 October 2026, at 20:20 UTC, Binance's own signing system approved 50 authorizations in a single Ethereum transaction.

At 20:24 the hot wallet known as Binance 14, the online wallet that pays out withdrawals, was empty. Three minutes and 24 seconds.

By the end of the evening about $6.3 billion had left the exchange across several blockchains. It is the largest crypto theft ever recorded, more than four times the previous record.

Many headlines called it a key theft. It wasn't.

Binance found no evidence that private keys leaked. Every transaction carried a valid signature produced by Binance's own system. The blockchains accepted them because, technically, nothing was wrong with them.

What the attacker took was access. A permission token on one engineer's laptop. A route into the service that asks the wallets to sign.

Bybit in February 2025 ($1.46 billion) and Bitget in September 2026 ($388 million) followed the same shape. Over $8 billion across three of the largest exchange losses ever.

No private key stolen in any of them.

This article puts the three cases in twelve years of crypto thefts, walks through each one and closes with the controls worth checking at any exchange or custodian you work with.

Twelve years, three ways to lose

Crypto thefts tend to get counted as one category. Look at what actually failed and the history splits into three phases.

The early years were about keys. Mt. Gox lost around 850,000 BTC by 2014, worth roughly $450 million at the time. Bitfinex lost 119,756 BTC in 2016.

Coincheck lost $530 million in NEM in January 2018 from a single hot wallet with no multi-signature protection, after employees' computers were infected through phishing emails.

The industry drew a simple lesson: protect the key. Cold storage and multi-signature became standard.

Then came the code years. In 2021 and 2022 the biggest losses came from bugs in smart contracts and bridges: Poly Network ($611 million), BNB Bridge ($570 million), Wormhole ($326 million).

Nobody needed a key. A flaw in the logic was enough. The answer was audits and bug bounties.

Ronin, in March 2022, sits on the hinge. Attackers took control of 5 of the 9 validator keys securing the bridge and withdrew $625 million.

The way in was a fake job offer: a senior engineer went through several rounds of interviews with a company that did not exist, then opened a PDF with the offer. The fifth key came from a signing permission one organization had granted another months earlier. Nobody had revoked it.

Since 2024 the largest thefts follow a third pattern. Keys stay where they are. Code works as designed. What fails is the layer in between: the people and systems allowed to ask for a signature.

At DMM Bitcoin in May 2024 an attacker posing as a recruiter on LinkedIn sent a "pre-employment test" to an employee of the exchange's wallet software provider. Weeks later a legitimate transaction request was manipulated and 4,502.9 BTC were gone, about $305 million.

2026 makes the shift measurable. In our ranking of the year's ten biggest hacks, compromised access and keys account for 84% of the $7.9 billion taken.

Binance alone weighs 80% of that total, so the honest reading needs a second look. Without Binance the picture is mixed: vendor and infrastructure failures at Bitget and Kelp DAO, a validation bug at Liquid Network.

Liquid is the useful contrast here. Eighteen days before Bitget, a flaw in the software that validates the network let an attacker mint unbacked tokens and withdraw about $320 million. Most of it, around 85%, came back.

Code bugs still happen. They are no longer where the largest sums go.

Explained simply: a key is not permission to use it

One distinction before the three cases.

Think of a vault with a time lock. Nobody can copy its key. The vault opens only for a correctly formed request. What it cannot do is look at who is standing in front of it.

An exchange wallet works the same way. The private key, or its MPC fragments (the key is split in pieces held in different places), sits in protected infrastructure.

Around it runs a signing service: software that receives a withdrawal, checks it against policy and asks the key to sign. Around that are the people, machines and vendors allowed to talk to the signing service.

Reach that outer ring and you don't need the key. You need the key to say yes.

Binance: four months of deception, 36 hours in plain sight

Binance's initial report of 3 November describes a campaign that started with people.

In June 2026 a wallet DevOps engineer, looking for a popular AI coding tool, installed a tampered package. Binance replaced his device and rotated his credentials, but the team that ran the wallet infrastructure was now mapped.

In September a second engineer was approached on a legitimate job platform. A fake recruiter, a fake HR representative on video who shared personal photos to build trust, a fake CEO in a chat app. At the end he was asked to run a coding test. He ran it on his work computer.

The malware replaced a component of a development tool and endpoint protection did not see it.

On 20 October the attacker used an internal AI tool to reach two more engineers. On one laptop he found a container deployment token. That token led to the internal signing service.

From there everything happened on-chain. Our own analysis of the transactions shows how visible it was.

On 28 October at 07:39 UTC the attacker's wallet received its first funding through Tornado Cash, a mixer. On 29 October at 15:59 it deployed the contract written to drain wallets and tested it at 16:30 on a fresh wallet.

At 20:20 Binance's system signed 50 EIP-7702 delegations, one for each nonce (the account's transaction counter), so that at least one would go through. EIP-7702 lets an account hand execution rights to a smart contract.

With the delegation live, a single call took 167,464 ETH, $1.6 billion in USDT and more than 50 other tokens.

On the other networks the funds left through ordinary transfers signed by the same system. Binance 16 sent 28,005 ETH to an address created 24 seconds earlier.

The delegation stayed live until 00:10 on 30 October. In those 3 hours and 50 minutes deposits kept arriving and the attacker took another 5,160 ETH.

Binance froze or intercepted about $2.85 billion and replenished 98% of the affected assets from its own funds. Its co-CEO Richard Teng wrote on 3 November that transaction-authorization mechanisms need new rules.

Count the steps that led there: a tampered package, a fake interview, one token on one laptop. Nowhere in that chain was a second person asked to confirm.

Bybit: what you see is not what you sign

On 21 February 2025 Bybit was moving ETH from its multi-signature cold wallet to a warm wallet. Routine work. The signers opened the transaction in Safe{Wallet}, the third-party interface Bybit used, saw the right address and approved.

What they could not see was that a developer's PC at the interface vendor had been compromised weeks earlier.

According to the vendor's forensic report, malicious JavaScript was added to the files served from its cloud on 19 February, written to activate only for Bybit's wallet. The hardware signing devices showed only a hash.

The instructions they signed replaced the cold wallet's logic with the attacker's. About $1.46 billion left within minutes. Two minutes after the theft the malicious code was removed.

Bybit's own infrastructure was never breached. Withdrawals never stopped and the ETH gap was closed in three days.

Recovery was another story: $42.89 million frozen in the first days. After one month 7.59% of the funds were no longer traceable.

The cold wallet was safe. The signing path was not.

Bitget: valid identities, forged commands

Bitget falls under infrastructure and vendors: its way in was a software flaw at a vendor.

On 31 August 2026 a zero-day vulnerability in a third-party security product, sitting inside the perimeter with high privileges, opened the way to the internal management systems. In September the attacker obtained high-privilege internal credentials.

In the eyes of the systems they were a valid identity. With them he moved to the wallet servers.

On 24 September at 18:31 UTC he sent two test transfers below threshold, 0.184 ETH and 193 TRX. No alert. From 18:58 to 20:09 seventeen large transfers followed, about $361 million.

The commands were written into the wallet system before the withdrawal records were created, so they skipped the risk checks.

At 19:05 reconciliation flagged the discrepancy and the system blocked user withdrawals across the platform. The forged commands were not user withdrawals: the last large transfer went out at 20:09.

Final figure: about $388 million from 12 hot and warm addresses on 11 chains. Cold wallets untouched, keys untouched, user balances covered by the protection fund. About $1.1 million frozen.

What limited the loss was real-time reconciliation, 34 minutes after the first transfer. The perimeter had not held.

Three roads, one shape

Side by side, the differences fade. The way in changes every time and two of the three came through a third party.

All three reached the same point: the system that turns a request into a signature. There, the controls checked whether the request was well formed. They did not check hard enough who was making it, or whether one access should be able to make it alone.

At Binance it was a deploy token that reached the signing service. At Bybit it was an interface the signers trusted because it was the interface. At Bitget it was an internal identity whose commands came in before the checks.

Each is a single point of failure in a process designed to have none.

Social engineering plus too much freedom for single accesses. That is a vulnerability of the system, not an isolated incident.

Multi-signature, MPC and cold storage solved the key problem. On their own they cannot solve the access problem, because they sign whatever the authorized layer sends them.

What this means for financial institutions

Most financial institutions in Europe will never run a hot wallet. Many are already connected to someone who does.

In the analysis we presented with Alfa Group at the CERTFin webinar on 16 September 2026, all 30 of the Italian institutions we looked at already send transfers to exchanges. 17 of 30 already have crypto or digital assets in house.

So the question for a compliance or risk team is concrete. When a counterparty can lose $6.3 billion in an evening without losing a key, what matters is how its access and signing model works.

That can be assessed. It belongs in counterparty due diligence next to capital and licensing. In the EU, DORA already treats ICT third-party risk as part of operational resilience for financial entities, crypto-asset service providers included.

There is also a timing lesson. Freezing works in the first hours, while funds are still in stablecoins or on exchanges that cooperate. Binance acted within minutes and froze about $2.85 billion. Bitget froze $1.1 million out of $388 million. The difference was speed.

The checklist above comes from our Lessons from the hacks section. Every control is tied to the case that showed why it matters. None of them needs blockchain expertise to ask about. A mature counterparty can answer each one.

Where Crypfy comes in: before signing

In all three cases the decisive moment was the signature.

After signing you chase. Before signing you prevent.

At Binance four signals were public before anything was signed: a contract already tested on a fresh wallet, funding from Tornado Cash, code written to drain wallets and a destination created 24 seconds earlier.

Transaction Simulation is the control built for that step. It simulates every custody transaction before signing, delegations and smart contracts included.

It reads the real effect: the destination, the assets that move, the permissions granted. It does this regardless of what the interface shows, which is exactly where Bybit's signers were misled.

The outcome comes back in real time with its reasons: allow, review or block. At Binance it is the check that would have flagged a wallet being delegated to a contract nobody had seen before.

Risk scoring covers the other half. A 0 to 100 score on the destination wallet with reason codes, on withdrawals as well as deposits and on small amounts too, so that Bitget's test transfers to new addresses would have counted as much as the large ones.

Both integrate via API into the anti-fraud engine a financial institution already runs, with Alfa Group.

The full breakdown of the three cases, the five patterns that keep coming back since 2018 and the complete list of controls are in our Lessons from the hacks section: crypfy.ai/insights/hack-lessons

How many of these controls does your counterparty already have?

Book a demo with our team: we show you the Binance, Bybit and Bitget cases read by the Crypfy console, then your own perimeter. crypfy.ai/contact