On 24 September 2026, Bitget lost about $387.5 million from its hot and warm wallets. The cold wallets were untouched and, according to the exchange and both forensic teams, no private keys were compromised. That last point matters most. The attackers didn't need the keys. They got the exchange's own withdrawal system to do the signing for them.

This is the same playbook that emptied Bybit in February 2025: leave the cryptography alone and attack the systems that decide what gets signed.

What happened

Bitget says the first unauthorized transfers began at 18:31 UTC on 24 September. Its reconciliation system flagged a mismatch at 19:05 UTC, and risk controls froze withdrawals across the platform. SlowMist's on-chain review puts the transfers between 02:31 and 05:23 (UTC+8) on 25 September, which is 18:31 to 21:23 UTC, so close to three hours across multiple chains. That means transfers kept landing for about two hours after Bitget says withdrawals were blocked. Neither report explains the gap yet.

The first estimate was $351.6 million. Bitget raised it to about $387.5 million after adding assets on Zcash and TRON. Its incident page now lists about $388 million across 12 hot and warm wallet addresses, 11 chains and 12 assets, including XRP, ETH, USDT, USDC, ZEC, BNB and AVAX.

Withdrawals came back in stages from 28 September and were fully restored on 2 October. Bitget says user balances were not affected and that its Protection Fund covers the loss.

The way in: the security tools themselves

Bitget commissioned two independent investigations, from SlowMist and Google Cloud's Mandiant. Both found that the attack started in two third-party security products, which they call Product A and Product B, or appliances A and B. Neither report names the vendor, and Bitget has declined to.

SlowMist's earliest malicious activity in the available logs dates to 31 August, more than three weeks before the theft. A zero-day in a service running on Product A let the attacker run a hidden script, read an environment variable holding a database password, and connect to the database. The same hidden-script activity showed up on other nodes on 23 and 25 September.

In the early hours of 25 September (UTC+8), the attacker logged into Product B's management platform using an internal employee's identity. Mandiant reports a web shell on appliance B, a command-and-control connection, and lateral movement from there to Bitget's production wallet job server, where malicious packages were deployed.

Forged withdrawals, real signatures

SlowMist recovered a custom withdrawal tool from files the attacker had deleted. The tool was built for Bitget's own withdrawal logic. It forged risk-control parameters, built withdrawal requests and called the normal withdrawal process. The wallet system treated them as routine and signed them.

Bitget's own description matches: stolen high-privilege credentials were used to write forged withdrawal commands straight into the wallet system, bypassing the risk-control checks that run before withdrawal records are created. Afterwards the attacker deleted traces. Even after the main transfers, logs show attempts to edit withdrawal records in the database and push fabricated BTC withdrawals, which failed.

One detail is worth noticing. Bitget's public page says the attackers "did not use common viruses or malware". Mandiant and SlowMist both describe malicious packages, a web shell and a custom tool. Both statements can be true depending on what you call "common", but defenders should read the forensic reports, not the summary.

Who did it

The working theory is North Korea, most likely the TraderTraitor cluster (also tracked as UNC4899). Bitget's CEO called DPRK involvement "very likely", citing IP addresses tied to VPN services used by a DPRK group. TRM Labs found on-chain overlaps with laundering wallets from Bybit and AFX Bridge, through a network it has only seen working for TraderTraitor. Elliptic rates the theft "highly likely" DPRK-linked.

That is a strong pattern, but it is laundering and method evidence, not proof from inside Bitget's network. Neither Mandiant nor SlowMist has published an attribution, and no government has formally attributed the attack.

The laundering looks familiar too. Funds were split into fresh wallets with round amounts, and most then sat still. The share that moved went through swap services, including THORChain, which refused Bitget's request to block attacker addresses. MistTrack reports automated CoW Protocol orders routed into Chainflip and on to Bitcoin, followed by CoinJoin.

Why this matters beyond Bitget

Three points stand out for anyone running custody.

Security products are part of your attack surface. The appliances meant to protect the network were the way in, and they usually run with broad network reach and high privileges.

Key security is not transaction security. Private keys stayed safe and almost $388 million still left. If any internal host can submit a withdrawal that your signer accepts, the signer protects the key, not the funds.

Dwell time beats detection. The attacker was inside for more than three weeks before moving money. Reconciliation flagged the theft 34 minutes after the first transfer, but only once funds were already leaving, and on-chain outflows continued after that.

The full breakdown, including the control points where this chain could have been broken, how to verify a withdrawal independently of the system that created it, and a hot-wallet hardening checklist, is part of the Advanced Security academy. The waitlist is open now.