Binance, banks and AI: the same power that builds can destroy

On 29 October 2026, Binance, the world's largest crypto exchange, suffered the biggest theft in the industry's history: around $6.3 billion by its own estimate. The way in, according to its first findings, was not broken cryptography or a superintelligence: it was an engineer's laptop. That same week, the largest US banks admitted they can no longer always tell whether the one logging into an account is the customer or an AI agent acting on their behalf. Both stories say the same thing: technology keeps getting more powerful, and whether it builds or destroys depends on how it is configured and who stays in control.

What happened at Binance

What is known as of 30 October, with figures still preliminary:

• Impact: Binance estimates around $6.3 billion in affected funds from its hot wallets. With partners it has frozen about $3 billion on-chain; Tether blocked more than $1.57 billion in USDT and Circle more than $800 million in USDC.
• Entry point: according to Binance, the attack may have started on the compromised laptop of a DevOps engineer, from which the attacker reached a service container able to submit withdrawals. The company says it has found no evidence that private keys were leaked and is still tracing the full attack path.
• Attribution: the analytics firm TRM Labs attributes the attack "with high confidence" to TraderTraitor, a group linked to North Korea. Binance has not confirmed it.
• Customers: Binance has suspended withdrawals while it rebuilds that service and says it will cover losses with its user protection fund and corporate funds.

One important nuance: as of today no source says the attackers used AI. They didn't need to. A machine with more permissions than it should have was enough.

What banks are seeing

Meanwhile, the financial sector has been raising alarms for a month:

• AI agents inside accounts. According to Axios (22 October), JPMorgan, Bank of America, Citi and Morgan Stanley have detected third-party AI assistants logging into customer accounts with their username and password, some doing things the customer never asked for. The bank sees a transaction with valid credentials and cannot tell whether the agent is authorised or has gone rogue. They are already restricting or blocking them. The banks say that, so far, no customer funds have been affected.
• Risk, multiplied. Jamie Dimon, CEO of JPMorgan, said on 6 October that cyber risk, already the bank's biggest, has gone up tenfold with the arrival of the most advanced AI models.
• Supervisors tighten the screws. According to Reuters, the European Central Bank gave euro-area banks until the end of October to present plans against AI-enabled cyber threats, and US supervisors are asking lenders whether they can audit their agents' decisions and shut them down in a controlled way.

The same power, two uses

AI is accelerating drug discovery, medical diagnosis and the productivity of companies of every size to levels that seemed unthinkable five years ago. In the right hands it is the best tool we have ever had. In the wrong ones, or simply badly configured, the same capability finds flaws faster, automates deception and acts without anyone having decided it should.

Binance shows that the weak link is usually the basics: a machine, a permission, a setting. The banks show that an agent holding the customer's keys without clear limits is a risk even if nobody built it with bad intentions. Neither is fixed by fearing AI. Both are fixed with discipline.

How we reduce risk at AizuaLabs

We design and maintain AI agents for businesses, so we ask ourselves these questions every day. This is what we do, not what we promise:

• Constant updates and patches. A watchdog checks known vulnerabilities in our websites, servers, machines and the code we maintain for clients every day. When a serious flaw appears, it gets patched, and we check before and after the change that everything still works.
• Guardrails in the code. No agent publishes, sends emails, charges money or changes sensitive data without a person's explicit approval. Our agents never claim to have done something they cannot do, and they always say they are an AI.
• Secure configuration by default. Connections with other platforms are signed and verified, code repositories are private and secrets do not live in the code. And every day we check from the outside, without credentials, that what should be closed is still closed.
• The client stays in control. They decide what each agent does, what they approve and what they don't, and can pause it at any time. An agent is a business tool, not someone holding the keys to the house.
• Offline as much as possible, online only where needed. What can be automated locally stays local; what has to be connected is locked down, signed and monitored. The less exposed a system is, the smaller the attack surface.

Five questions for your business this week

1. Which machines or people can move money or data without a second approval?
2. When were your websites, servers and software last updated?
3. Is any AI assistant using your passwords to log into your bank, email or ERP?
4. Can you switch off an automation within a minute if something goes wrong?
5. Does anyone check from the outside that what should be closed is still closed?

If any answer makes you uncomfortable, now is a good time to review it. In the free AizuaLabs audit we go through how to use AI in your business with control always on your side.

Sources: Binance statements reported by CyberInsider, Al Jazeera and TechCrunch; Axios via Infobae; Jamie Dimon interview reported by Investing.com; Reuters via Insurance Journal. Figures as of 30 October 2026, subject to change as the investigation continues.

Frequently asked questions