Superintelligence may still be years away. AI agents inside bank accounts aren’t.
The day after the Binance hack that took place October 29th, Elon Musk did what Elon Musk does best. He posted about it.
On X, he said: “Super intelligence agents will at some point do this to gain currency for themselves, not a human,”
He was referring to a clever move made by the hacker during the attack. Before issuers had the chance to freeze any stablecoins, hackers moved around $60m in stolen USDT into a crypto lending platform. Then, used it as collateral to borrow around $44.6m worth of Ether.
Reason being, issuers can freeze USDT (and that’s exactly what Tether and Circle did to a large chunk). They cannot, however, do the same to Ether in the same way. So by swapping one for the other, the attacker found a way to move money beyond reach.
Musk’s point was that if a human can spot and exploit a weakness like this, an intelligent enough AI agent will eventually be able to do so too, and it won’t necessarily be working for a human but instead acting in its own interests.
The more I looked into it, the more I realised we might be a lot closer to Musk’s prediction becoming reality than I’d like to think. Here’s why…
Washington already says it’s here
Back in September, the White House signed an executive order telling federal agencies to stop using the term ‘artificial intelligence’ and, instead, start calling it Super Intelligence.
The order defines Super Intelligence using the same legal definition as artificial intelligence. So this is essentially just semantics, not a declaration that superintelligence has arrived. It does, however, mean that the US government has begun using Musk’s terminology in its own guides and documents.
So what about the technology itself? Are we actually close?
Last month, Deutsche Bank‘s Research Institute looked at a theory known as the San Francisco Consensus. It’s the belief, one shared by many over at Silicon Valley, that AI will eventually start accelerating its own development and potentially reaching superintelligence by 2030.
Researchers haven’t ruled this possibility out, though they have pointed to some obstacles, including scientific uncertainty and the enormous amount of electricity and computing power that would be required.
But I think we’re getting distracted by the end goal here. Because you don’t need a machine that’s smarter than every human to create serious problems for banks and other financial institutions. All you need is an AI agent that can spot a weakness, then act on it by moving money before anyone notices. That technology exists today.
The bots are already in the accounts
Almost a year ago, Anthropic published research showing that AI agents could find and exploit weaknesses in real cryptocurrency smart contracts in a simulation.
In just 12 months, the share of recent attacks they could successfully carry out rose from 2% to more than half. Finding weaknesses in a contract cost as little as $1.22.
I want to be clear here about what this means. This isn’t a prediction made by a tech billionaire on what AI might be capable of in the future. This research was proof of what these systems can actually do, right now, in a controlled environment.
Then, on 22 October, Axios reported that some of America’s biggest banks were discovering unapproved AI agents inside customer accounts. I’ve recently gotten really into using AI-bots myself, using them, both for streamlining my work, and automating my personal life. There was a time whereby the bot platform I use asked if I wanted it to access my banking information. I, of course, did not allow it access. That doesn’t seem to be the case for everyone using these bots.
Axios found, customers had willingly given AI bots access to their own banking information. Bots had logged in using customers’ own passwords. Some even appeared to be carrying out transactions or other actions that customers hadn’t actually requested.
Banks moved quickly to restrict access. Morgan Stanley began blocking third-party AI agents, although it said it hadn’t found any rogue agents itself. JPMorgan, Bank of America and Citi also tightened restrictions on access to their systems.
Freezes work, just not fast enough
Back to the Binance hack. The industry’s response, to its credit, was really strong. I wrote about this recently because I was really impressed by how quickly stablecoin issuers stepped in to freeze the stolen funds. Without those freezes, the losses could have been considerably worse.
However, while the first freeze came in under an hour, it was still after attackers made the Aave trade and a large chunk of the money had moved beyond reach. The FBI has since blamed North Korean hackers for the attack. People, in other words, using automated tools to move quickly and exploit opportunities as they appeared.
That’s what worries me because, if a human-operated team using AI can get that far in 40 minutes, how much could an AI agent achieve in the same amount of time?
Regulators are asking the right question
The most useful thing I’ve read on this came from America’s national bank regulator, the OCC. On November 4th, it published a framework telling its examiners how to judge banks’ use of super intelligence.
Its big idea is that the risk shouldn’t be based on how clever an AI is, but what it’s allowed to do. A chatbot that drafts internal memos sits in the lowest risk tier, for instance. A fully autonomous agent that can send payments, in contrast, sits in the highest however well it does in testing.
The OCC makes a point I think everyone in payments should take seriously. An agent doesn’t have to be hacked to cause damage. It can be running perfectly and still do something it should never have been allowed to do.
In a survey of 230 US banking professionals this spring, 72% said their institution is not prepared to shut down an AI system that goes wrong, or to report the failure to regulators.
Here’s what I’d do to prepare for these occasions:
- Know every agent on your rails: That includes your own, as well as your vendors’ and customers’. Know what each one can touch.
- Give agents an identity and limit: Every agent that can move money should be tied to a verified owner, with spending limits you can change or revoke.
- Build brakes that don’t wait for a human: Limits that stop unusual payments automatically, with a person signing off anything big that can’t be reversed.
- Test your kill switch: On a system that looks like it’s working fine.
Agree the response in advance: Banks, exchanges and issuers. All should know who calls who before something happens, not during.
At some point
Musk said superintelligent agents would do this “at some point” and I personally think he may be right. But, I also don’t think that this matters as much as what’s already happening today.
Because AI bots are already logging into banks. They’re already using tools to find their weaknesses, tools that can cost as little as a dollar.
Our best response to a $6.3bn theft still started with someone noticing and reporting it quickly.
The real test of financial security won’t be how much we freeze after the fact. It’ll be how little we let leave in the first place.
The post Superintelligence may still be years away. AI agents inside bank accounts aren’t. appeared first on The Fintech Times.