Fine-grained controls could have prevented the Binance hack
Binance lost $6.3 billion without losing a key: its signing service signed whatever a compromised insider asked. The missing layer is policy enforced in the signer.
TL;DR
- Binance lost about $6.3 billion on 29 October without losing a single private key. The attackers took over the service that asks the key to sign, and it signed whatever they asked.
- Key security only answers one question: can an attacker get the key? MPC, HSMs, enclaves and multisigs all answer it well, but in this case no key was needed for the theft.
- What was missing is control over what a valid signature may do: fine-grained, per wallet, enforced inside the signer itself.
- With that layer in place, 29 October ends as fifty refusals and an alert instead of a $6.3 billion loss. That layer is what we built Pakt to be.
Lessons learned from the hack
- Coarse controls do not survive a compromised insider. The attackers were standing in the application tier with valid credentials. To the signer, whoever holds the credential is you.
- A hot wallet has a job, and the key should only be able to do that job. A payout wallet pays registered addresses. A market-making wallet posts orders on a few venues within a daily size. A solver wallet fills quotes. The job is narrow, but the key's authority is total. That gap creates an opportunity to attack.
- Fine-grained controls should be written in the vocabulary of the job:
- Allowed payload types, so a payout wallet physically cannot sign a delegation or an arbitrary message.
- Outflow per hour, so a drain that should take days cannot happen in three minutes.
- Rules on markets, leverage, and equity floor.
- One policy across transfers, trading and lending, so a limit in one product cannot be routed around through another.
- And a check against the account's projected state, because the dangerous orders are often individually small.
- Enforce where the signature is made, and fail closed. The policy has to live inside hardware the attacker cannot reach with stolen credentials. If a rule fails, no signature exists, and there is nothing to intercept.
- Be able to show it. A config file gives you a config file. A proof-gated signer gives you, for every signature, a statement anyone can verify that it stayed inside the rules, and for every refusal, a record of why.
What happened on 29 October
Binance's own findings, published on 3 November, read like a spy novel with a bad ending.
In June, a wallet DevOps engineer downloaded a malicious installer while looking for an AI tool. In September, a second engineer went through a fake recruitment process, complete with an HR round and a "CEO" interview, and was handed a take-home assignment that was malware. In October, the attackers used those machines and an internal AI tool to reach a third engineer, borrowed a container-deployment token, deployed a container, and from inside it reached the internal wallet signing service. In Binance's words, they used it "to initiate transactions outside the normal withdrawal process."
The drain began at 20:05 UTC on the XRP Ledger. At 20:20, the attackers submitted fifty EIP-7702 authorizations, all pre-signed by Binance's busiest Ethereum hot wallet, each handing the wallet to a contract the attacker had deployed that afternoon. There was one for each of fifty consecutive nonces, so whichever matched would go through, and one did. Three minutes and twenty-four seconds later, the wallet was empty.
The same wallet was then used to pull 823 million USDC out of a Binance smart account it could drain. On Bitcoin, Tron, XRP Ledger, Solana, and five more chains the drain was plainer: ordinary signed transfers, 18,270 BTC among them. Tether, Circle, Ripple and the Arbitrum Security Council froze roughly $2.8 billion. The rest is gone.
Binance's remediation list is what you would expect: rebuilt the wallet system, replaced devices, rotated credentials, added anomaly detection, tightened endpoint controls. All of it keeps attackers further from the signer. None of it, as published, changes what the signer will agree to sign. That is the layer that was missing.
The same night, with fine-grained controls
Now run it again with one change: Pakt sits in front of the signer.
Each hot wallet has a policy, set with a cold key that lives in a multisig or at a custodian and is touched only at setup and recovery. The hot key is generated inside a trusted execution environment and never leaves it. Every proposed action, from a script, a bot, an agent or a human, is checked inside the enclave against the policy and against the account's projected state before anything is signed.
The attacker has the container, the signing-service credential and the fifty delegation authorizations. The payout wallet's policy does not include EIP-7702 delegation. Fifty refusals.
They fall back to plain transfers. The policy caps hourly outflow and restricts destinations to registered addresses. Refused. They find the smart-account path and ask for 823 million USDC. Not allowed. Refused.
Total loss: zero, a very loud alert at 20:05, and a cryptographic receipt for every refused request that the security team can hand to the FBI.
Who this is for
Exchanges, obviously. But also market makers, solvers, bridges, OTC desks, payment processors, treasury teams and any trading firm that keeps operational keys hot because the job requires it. If a server can sign on your behalf without a human in the loop, you run a hot wallet, and the Binance findings describe your threat model whether or not anyone has ever offered you a fake job.
Pakt is that control layer: fine-grained policies enforced at signing, inside a TEE, with a zero-knowledge proof for every decision.
We are running custom pilots with custodians and trading firms now, each built to the wallet's actual job. If you operate hot wallets and want to know what yours would look like with controls that hold even when the credentials don't, talk to us.