How a Backend Compromise Drained $387.5M From Bitget Without Touching a Private Key

At 18:31 UTC on September 24, 2026, Bitget's security systems detected unauthorized transfers moving assets out of a portion of the exchange's hot and warm wallets. By the time the outflow was contained, roughly $387.5M in ETH, XRP, BNB, ZEC, USD0, XAUT, AVAX, TRX, USDT and USDC had crossed Ethereum and other EVM networks, the XRP Ledger, Zcash and TRON. Nothing about the failure involved a stolen private key: Bitget's investigation ruled key compromise out entirely, and its cold wallets were untouched throughout.

What the attacker compromised instead was the backend system that assembles and authorizes the exchange's transactions — the machinery that decides what a key is asked to sign — and used it to spoof transaction data and trigger the authorization process itself. That is the largest loss in this record, executed against the layer every previous edition assumed was working.

1. The Incident in Six Lines

2. Hot, Warm, Cold: What the Three Tiers Are For

Every large exchange runs a three-tier wallet architecture, and the tiers exist because of a trade-off no operator can escape: keys that can pay out quickly are keys an attacker wants, and keys an attacker cannot reach cannot pay out quickly. Hot wallets sit online and handle immediate withdrawals; warm wallets sit between them and cold storage, batching larger flows; cold wallets keep their keys offline and are the reserve. Bitget's three-tier design is exactly why the incident stopped where it did — the breach touched a portion of the hot and warm layers, and the cold reserve, where most platform assets sit, was never reached.

But the tiers answer only half the question. Between the user's withdrawal request and the key that signs it sits a pipeline: services that validate the request, assemble the transaction, apply policy checks, and hand the payload to a signing system. That pipeline is software running on infrastructure, and in this incident it was the compromised layer. Bitget's account is precise about the mechanism — the attacker used the compromised backend to spoof transaction data and trigger the authorization process. The keys did their job: they signed what the pipeline told them to sign. The failure was that the pipeline was made to lie.

3. Timeline

4. Where It Broke: The Authorization Pipeline

The specificity of Bitget's language is the most instructive part of the disclosure. The exchange did not describe a leaked key, a phished employee credential or a smart-contract bug; it described transaction data being spoofed and an authorization process being triggered — by the attacker, through infrastructure the attacker controlled. Private-key compromise was ruled out explicitly. In the terms this record has used for three weeks, the distance between the key and the payout was crossed on the pipeline side: the system that decides what a key is asked to sign was made to request the wrong signatures, and the signing layer correctly approved them.

The closest precedent remains the Bybit theft of February 2025 — $1.4B, the largest on record — where attackers spoofed a signing screen to hijack a routine cold-wallet transfer. The two incidents bracket the same lesson from opposite ends: Bybit's users were told a lie about what was being signed; Bitget's backend was made to request signatures for transactions that should never have been assembled. In both cases the key material was never wrong and never left the operator's control. What failed was the trust that the thing asking the key to sign is honest — which means an exchange's security boundary is not its keys but its whole transaction pipeline, and a user evaluating an exchange is evaluating software they cannot see.

5. Who Was Actually Exposed

The answer divides cleanly, and not along the lines the headline suggests. Users whose assets sat in their own wallets — on exchanges elsewhere, in self-custody, on-chain — were never exposed, whatever chains they used. Users with balances on Bitget lost no principal: the exchange states account balances were accurate throughout, deposits and trading continued, and the loss falls within a User Protection Fund published at over $464M. What those users actually lost was time and control. All withdrawals froze for four days; restoration runs on a phased schedule of the exchange's announcement, asset class by asset class, from September 28. The exposure condition, stated plainly: anyone whose money was inside the trust boundary when it failed experienced a freeze the trust boundary's owner controlled — regardless of whether the breach touched their assets at all.

Two narrower exposures deserve their own lines. First, holders of the affected assets: the confirmed list includes TRX and USDT, and the revised estimate explicitly added Zcash and TRON transactions — a reminder that multi-chain incidents reach every chain a venue touches, TRON included, and that issuer-freezable assets were the attacker's first disposal problem rather than the victim's. Second, the on-chain observers: early independent estimates ran $174–183M, roughly half the final figure, and one researcher's analysis even referenced a cold wallet among affected addresses while the exchange maintained cold storage was fully secure — a discrepancy that comes down to how third parties label the warm tier. When the labeling is contested, the number is contested; the honest posture toward any incident figure in its first days is that it is a floor moving in one direction.

6. The Rest of the Week on the Same Boundary

Four other losses landed in the same seven days, and none shared a mechanism with Bitget. What they shared was the property: value moved on the strength of a system's own say-so, and the say-so was wrong.

A governance proposal on Neutron cost about $20,199 in voting power, passed with eleven Update Admin actions, and lawfully redirected the administrators of contracts used by the Astroport exchange and the Drop liquid-staking protocol — roughly $9.3M per SlowMist's tracker, with Neutron's rate limits and halt containing much of it and Cosmos Hub validators sweeping the attacker's remaining 1,227,121 ATOM, about $2.2M, into a multisig on a coordinated restart. Nothing in either application's own security posture could object, because the takeover was authorized one level down, by the chain itself. A block-validation flaw on Meter let an attacker mint unbacked MTR and MTRG on September 23 and sell roughly $2.3M into PancakeSwap, crashing both tokens by 75% or more — the first incident in this record where consensus rules were the broken layer. The casino Duelbits lost approximately $7M to a suspected hot-key compromise spanning five chains, its estimate climbing from $4.3M to $5.9M to the co-founder's own ~$7M as investigators found more addresses. And a single leaked private key drained more than 145 Ethereum addresses plus TRON and Bitcoin holdings for over $4.3M, still active at the alert's publication — the standing case that individual key hygiene has not stopped being the first line of defense even in the week the pipeline became the target.

7. What to Do With This

• The exit: rehearse a withdrawal from every venue you depend on — small, quarterly, documented. Bitget users lost no principal and four days of access; the difference between inconvenience and crisis was whether the exit had ever been tested.
• The sizing rule: hold on any custodial venue what you can afford to have frozen, not what you can afford to lose. Every incident this week ended in a pause the user did not schedule, and the asset kept its price while access was gone.
• The keys: assume any key that has touched software is discoverable. The week's 145-address drain is the running argument for generating significant holdings on a clean device and leaving them there.
• The trust boundary question: before holding value on any platform, ask what lies between your withdrawal request and the signing key — and whether the operator publishes anything about that pipeline. The keys being secure is now demonstrably not the same as the wallet being safe.
• The scam surface: after any incident of this size, restoration notices arrive by message and social media that are not from the venue. Bitget's own channels carried the phased schedule; everything else is bait. Verify the channel before you verify the news.

8. What to Watch

Three questions decide what this incident means beyond Bitget. First, the root cause: the exchange has promised a full report and has deliberately not speculated on the attack vector; whether the report names a weakness the rest of the industry shares — a backend pattern, an authorization design, a vendor component — determines whether $387.5M bought the sector a fix or only bought Bitget one. Second, the fund arithmetic: drawn at the revised figure, roughly $77M of the published $464M headroom remains, and every exchange with an insurance fund is now being asked what its own coverage would look like after a comparable event. Third, the pace of restoration: a four-day freeze absorbed calmly in a rising market is the best-case reading; the same freeze in a falling one is the stress test nobody chose. The deeper question underneath all three is whether the industry's security spending follows the losses — which, for three consecutive weeks, have been moving from the keys to the systems that use them.

Disclaimer: This content is for informational purposes only and does not constitute financial, legal or security advice. No checklist or tool eliminates risk entirely; self-custody means you are responsible for your own keys.