Are AI crypto trading bots safe? $387.5 million left Bitget without a stolen key

Are AI crypto trading bots safe? Bitget lost $387.5M without a stolen key. Four questions show how much your bot or AI agent can move.

Published 10.15.26 · updated with the SlowMist and Mandiant findings Bitget published on 30 September

Whether a crypto trading bot or AI trading agent is safe comes down to four things: who holds the key that can move funds, how much one bad instruction can move, whether anyone reads each transaction, and which approvals are still sitting on your wallet. In September an attacker moved $387.5 million out of Bitget without stealing a private key. The withdrawal system did what it was built to do: it approved instructions that matched its rules, after the attacker had spent three and a half weeks inside learning them.

Custody and cryptography both held up. What failed was autosigning: funds moved because code evaluated rules, and no person checked what the transfers were for. Every AI agent that executes trades depends on the same arrangement, since the product is that you do not have to be there when it trades. We ran the four questions on Reticle too; the result is near the end.

What is autosigning?

Autosigning is any arrangement in which software produces a valid cryptographic signature for a value-moving transaction without a human reviewing that specific transaction first.

The phrase covers more setups than most people picture. An exchange's internal system that signs withdrawals after automated checks is autosigning. So is a trading bot holding a hot key. A token allowance you granted once and never revoked is deferred autosigning: you pre-authorized a contract to move your funds under conditions you no longer watch. A session key with a spending cap is autosigning with a ceiling.

In all of these, the decision to move funds is made at that moment by code checking rules.

How did $387.5 million leave Bitget without a stolen key?

On 24 September 2026, Bitget detected unauthorized transfers out of its hot and warm wallets (Bitget's account). The figure settled at $387.5 million across 11 blockchains, revised up from an initial estimate of $351.6 million. Cold wallets were not affected. Withdrawals were suspended and then restored in stages, completing on 2 October. Bitget absorbed the loss rather than passing it to users. Recovery has been small: roughly $1.1 million frozen by Circle, Tether and NEAR Intents, about 0.3% of what was taken. (BleepingComputer, CNBC)

No private key was stolen. Bitget's chief executive said so on the record after the first reports had settled on key theft (CoinDesk). Per the SlowMist and Mandiant investigations that Bitget published on 30 September, the attackers exploited a zero-day vulnerability in a third-party security product running inside Bitget's environment, ran a hidden script to read a database password out of a service's environment variables, and established access from there. SlowMist dates the earliest malicious activity in the available logs to 31 August 2026. The transfers happened on 24 September. (The Hacker News, citing SlowMist and Mandiant)

Until then the foothold was a hidden script on a single node. Mandiant found that around 24 September the attackers gained access to the security appliances, deployed a web shell on them, established a command-and-control channel, and moved laterally to Bitget's production wallet job server, where they deployed malicious packages. SlowMist recovered a deleted tool from the environment and described it as highly tailored to the wallet system's withdrawal logic. Bitget's own account is that the attackers obtained high-level internal credentials and used them to issue fraudulent withdrawal commands, producing "abnormal transfers that bypassed existing risk controls."

In sequence: in through the security software, three weeks of learning how the withdrawal pipeline decides what to approve, a tool built to speak that pipeline's language, and then a request for the money.

Nothing cryptographic failed, so no cryptographic fix exists. The pipeline enforced every control it had, correctly, on instructions that no human read.

Bitget's chief executive said IP behaviour patterns and on-chain analysis point to North Korean actors (The Record), and Elliptic (via Bloomberg, republished by Insurance Journal) and TRM Labs have reported overlaps with laundering wallets from earlier thefts. The argument below does not depend on the attribution.

Why was the way in a security product?

The entry point was software Bitget had bought in order to be safer. A security appliance has to sit inside the trust boundary to do its job, and that position is also why attackers go after it. Bitget has since notified the vendor and disabled the affected functionality pending a fix.

Carry that over to anything you let near your own funds. The agent's own code is one question. The RPC endpoints it calls, the price feeds it trusts, the libraries it imports and the extension it signs through are all part of the same surface, and you have probably audited none of them.

Why does this matter for your AI trading bot?

After an exchange breach, most advice is about custody: move your coins off and hold your own keys. Custody matters, but Bitget showed something else. An automated approval pipeline can be reverse-engineered by anyone who gets three weeks inside it, and then fed instructions it is designed to approve.

Much of what the crypto industry built in 2026 works the same way: agents that read markets and execute, bots that rebalance overnight, assistants that turn a sentence into a chain of on-chain actions. They exist so that you do not have to be awake when they act, so most of them have to autosign. What you can find out is how much authority the autosigning has, and how much is lost when something feeds it a well-formed lie.

A ruleset becomes attackable because it is consistent. Automated approval has to be deterministic to be useful, which also means it can be learned. Bitget's attackers did not defeat its rules; they spent three weeks reading them.

This is the first time that failure has happened at institutional scale in public. Our guess is that the next case is more likely to be a consumer agent than an exchange.

What are the four kinds of signing authority?

Custodial versus non-custodial is too coarse. There are at least four arrangements, and each fails differently.

ArrangementWho can move fundsFailure modeBlast radius
Full custodyThe platform, under its own controlsPlatform internal compromise, of which Bitget is the reference caseEverything the platform holds, pooled
Delegated keysA bot or agent holding a hot key on your behalfKey extraction, or the agent being fed bad instructionsEverything that key controls
Session keys and allowancesA contract or agent, up to a pre-authorized limitForgotten standing approvals; a compromised contract draining within its capWhatever you pre-approved, which is often unlimited
Per-transaction human signatureOnly you, once per transactionPhishing, blind signing, social engineeringOne transaction, if you read it

The bottom row limits the loss to one transaction, provided you read what you sign. None of the rows earns the word "safe."

How much can your trading bot move? Four questions

These apply to a trading bot, an AI trading agent, or any platform that can touch your funds.

Who holds the key that can move funds?

Asking "who custodies the assets" usually gets a marketing answer. The specific question is which party holds the signing material that can produce a valid transfer.

If that party is anyone other than you, their internal security is now your security, including their backend systems, their deployment pipeline, their vendors and their worst day. Bitget's users were not compromised. A security appliance in Bitget's environment was, and that was enough.

Then ask whose code can reach the thing that signs. The key holder is one answer. The list of software with a path to the key holder is longer, and in Bitget's case it included a product bought to reduce risk.

How much can a single forged instruction move?

This answer sets what a breach costs you. An agent that can move your entire balance has an unbounded blast radius. An agent capped at a fixed amount per transaction, per day or per counterparty has a bounded one. In an attack that works by making bad instructions look good, every other control has already been satisfied, and the cap is the one that still holds.

Ask for the number. Bitget shows what its absence costs: the pooled hot and warm wallet infrastructure had no per-instruction ceiling that stopped the sequence, $387.5 million left across 11 chains, and roughly $1.1 million of it has been frozen.

Does a human review what the transaction does, or only its format?

Automated validation checks whether a transaction is well-formed: correct fields, valid signatures, consistent state, within the rules. It cannot check whether the transaction is one you wanted, because wanting is not a property of the bytes. Only a person who knows what they intended can make that call.

So "you sign every transaction" is a weaker claim than it sounds. Someone who approves a wallet prompt without reading it is a slower autosigner. A product that puts a human in the loop and then designs the prompt to be clicked through keeps the liability and loses the benefit.

What standing approvals already exist on your wallet?

Token approvals persist. You granted an allowance to a DEX in 2024, the interface asked for an unlimited amount because that saves gas later, and that allowance is still live. If that contract is ever compromised, the attacker does not need your key, because you already signed.

The custody debate tends to skip this one: your keys can be perfectly safe while an old permission moves your funds.

Is self-custody safe?

Holding your own keys swaps one set of risks for another. You remove exposure to a platform's internal compromise and take on phishing, malicious signature requests, wallet-draining sites, address poisoning and your own attention on a bad night. Losses in that category are not rare, and unlike Bitget's users, you have no protection fund.

Ledger, which sells self-custody hardware, publishes a Ledger Academy piece called "Addressing Self-Custody's Greatest Misconceptions." A vendor writing that about its own category is a fair sign the misconceptions are common.

Neither model wins outright. You should at least be able to name which risks you have taken on, and the four questions are a way to find out.

How does Reticle answer the four questions?

Reticle is Antalpha's Web3-native AI agent (NASDAQ: ANTA). It reads markets across crypto, prediction markets and US stocks, and drafts transactions that the user signs from their own wallet.

We built Reticle around a signature gate. Antalpha's documentation describes the architecture as a five-step flow where step 4 is the gate: the transaction is drafted, and the user signs. The documentation says it plainly: control of a user's assets "always stops at step 4 — in your own wallet."

Run the four questions against it. The user holds the key, because Reticle connects to the user's own wallet and never stores one. The blast radius of a single forged instruction is one transaction, and only if the user signs it. The review happens at the signature gate, where approval is a separate human act the architecture cannot skip. Standing approvals are the one place Reticle acts on your behalf instead of waiting: it runs a wallet approval checkup across chains, surfaces risky allowances and revokes them with a single signature, and it scans a token contract before a swap and blocks the trade if the contract is dangerous.

Reticle's other features sit in different places on that ladder. Sentinel watches markets and sends alerts; it never places an order. Vibe Trading, Reticle's AI trading bot, executes by rule, which makes it a form of autosigning; today it runs only in paper trading, moves no real money, and live trading is not yet available. When live trading opens, the same four questions apply to it.

The third question is only as strong as the user makes it. Reticle puts the review in the user's hands, and a user who approves without reading has given that protection away.

Reticle is an intelligence layer, and it is not a security product. It reports the freshness and origin of the data behind every answer through a badge with states for verified, delayed, partial and insufficient, and it declines to produce a number when the data does not support one. The full source list with each source's update cadence is published at How Reticle Works. That covers the analysis; it is not an audit of the chains you transact on.

You should not have to take our word for the approvals part either. Any block explorer will list the allowances sitting on your address. Reticle's checkup should agree with it, and if the two ever disagree, trust the explorer.

What can you do in the next ten minutes?

None of this requires a tool.

Run an approval checkup on your wallet and revoke every allowance you do not currently need. Most people find at least one unlimited approval they had forgotten about, granted to an interface they no longer use. Then, for every bot, agent or platform that can touch your funds, write down the answer to the second question. If you cannot find a number, treat the exposure as your full balance; a missing cap is an answer in itself. Finally, the next time a wallet prompt appears, read the destination and the amount before approving. That reading is what per-transaction signing is for.

Bitget's controls all worked as specified, for an attacker who had taken the time to learn what they would accept. A person who knows what they meant to sign is the one check that studying the rules does not get past.

Sources

Frequently asked questions

Are AI crypto trading bots safe?

Safety depends on the signing architecture more than on the quality of the model. Ask who holds the key that can move funds, how much a single instruction can move, whether a human reviews what each transaction does, and what standing approvals already exist.

How should I set API key permissions for a trading bot?

Grant trading permission only, never withdrawal or transfer. Most exchanges also let you restrict a key to an IP allowlist, so it only works from servers you name, and let you lock withdrawals to an address allowlist; on OKX, for example, withdrawals then only go to addresses saved in your address book. If the key leaks, whoever holds it can place orders but cannot move funds out. Delete keys you no longer use.

What is autosigning in crypto?

Autosigning is any arrangement in which software produces a valid signature for a value-moving transaction without a human reviewing that specific transaction first, including exchange backend signing pipelines, bots holding hot keys, and standing token allowances.

Was the Bitget hack caused by stolen private keys?

No. Investigations by SlowMist and Mandiant found that attackers exploited a zero-day vulnerability in a third-party security product inside Bitget's environment, moved laterally to the production wallet job server, and used high-level internal credentials to issue fraudulent withdrawal commands that bypassed existing risk controls. No private key was stolen.

How did attackers get into Bitget?

According to SlowMist and Mandiant, the attackers exploited a zero-day vulnerability in a third-party security product running in Bitget's environment, with the earliest malicious activity dated to 31 August 2026, then moved laterally to Bitget's production wallet job server before transferring funds on 24 September 2026.

Does self-custody make crypto safe?

No. Self-custody removes exposure to a platform's internal compromise and adds exposure to phishing, malicious signature requests, wallet drainers, and forgotten token approvals. You still carry risk; it is a different set of risks.

What is the most overlooked wallet security risk?

Standing token approvals. An unlimited allowance granted once to a contract remains active indefinitely, and a later compromise of that contract lets an attacker move funds without ever touching a private key.

Can Reticle move funds without the user's signature?

No. In the Reticle app, Reticle connects to the user's own self-custodial wallet, never stores a private key, and requires the user to sign every transaction from their own wallet. Sentinel only sends alerts and never places orders. Vibe Trading executes by rule but runs only in paper trading today, so it moves no real money; live trading is not yet available.


Reticle is the Web3-native AI agent from Antalpha (NASDAQ: ANTA), covering crypto, prediction markets and US stocks: conclusion-first analysis with named sources and freshness badges, and Sentinel mode for round-the-clock monitoring. Vibe Trading, Reticle's AI trading bot, turns a one-sentence trading idea into a rule-based bot that you backtest on historical data and then run in paper trading; live trading for Vibe Trading is not yet available. Backtest results are simulated on historical data and do not indicate future returns. Trades in the Reticle app are signed by you in your own wallet by default; Reticle never holds your funds and takes no cut of principal or profits; a transparent routing fee applies on the execution path. Nothing here is investment advice or a price prediction. Official channels: antalpha.com, the Reticle app on the App Store and Google Play, x.com/antalpha_ai.

Related reading: How Polymarket odds work: a five-step check for whether an edge is real.

Reticle

Everything these articles mention lives in the Reticle app

Reticle checks the token approvals, suspicious contracts and standing signatures on your wallet and tells you which gate is still open.

Want to see where your wallet approvals stand? Open Reticle and run a wallet health check
Subscribe to InsightsRSSXDiscord

Author & sources

Antalpha-Evan · Research Team · Trading Safety Researcher

Sources for this article ↑