A recent incident in Australia shows that giving AI systems goals instead of strict step-by-step instructions can get a lot more complicated when those systems are connected to something valuable.
An Australian man asked his Claude-powered OpenClaw AI assistant to book a popular gym class.
However, the agent discovered a vulnerability that allowed it to book classes much further in advance than the gym intended. It figured out that the gym’s API lacked authorization safeguards for canceling others’ bookings. The agent used the flaw to cancel an existing booking and move its user up the waitlist.
The important thing to mention here is that the AI agent wasn’t instructed to do that, but rather saw the situation as a technically possible way to get closer to achieving its goal.
A Warning for All, Including Crypto #
Now, if we replace a gym booking with a crypto wallet, the results could be much more damaging.
In the event an AI agent is given access to a wallet and told to maximize returns or find the best DeFi deals, how do we define what it’s actually allowed to do?
In theory, an advanced enough AI could stumble across a weak point in an API, interact with a smart contract, approve token spending, shuffle funds between platforms, or make a trade that technically fulfills its stated objective but does something the user never wanted.
This doesn’t mean current AI agents are destined to go rogue. Still, the gym booking incident shows why this kind of risk is worth paying attention to.
Who’s Responsible When an AI Agent Makes an Unauthorized Decision? #
The responsibility for unauthorized AI agent actions is hard to determine because there are no laws against such instances yet. It could also largely depend on intent. For example, if someone knowingly grants the AI agent wide-ranging authority, they could be held responsible for what it ultimately does.
Additionally, an AI developer may face scrutiny if the system was designed or promoted with insufficient care. Similarly, a wallet provider might face questions if they didn’t have strong enough safeguards, and DeFi protocols or API operators could also share blame if their own security flaws allowed the AI’s action to happen.
As of today, there isn’t a universal rule making the AI itself responsible, but if incidents like this repeat, there may be some changes.
**Related: **AI Agents Could Transform How Money Moves Across the Internet