AI has graduated from hacking companies to stealing from crypto wallets

A couple years ago, reading headlines like “artificial intelligence resorts to hacking to complete a task” might have sounded like something out of science fiction. In the present, however, that’s just the world we live in, and multiple stories like this have already emerged over the past few months. It looks like AI’s rogue era is only continuing, as we learn about an incident where OpenAI caught an internal research model committing some minor theft in order to secure the funds it needed.

In its incident report, OpenAI details what it learned from the model’s behavior this past June. The model was assigned a cryptocurrency-related task, working on Ethereum’s test network, Sepolia. While the model wasn’t given access to its own wallet or any funds with which to perform actions, it should have been able to acquire test-network tokens to let it continue — except, here’s where things took a turn.

Instead of figuring out how to deploy a test contract, the model decided that it needed to secure some resources first, and set out to do just that. It managed to compute a list of common developer recovery phrases, and from those it was able to identify an account with a balance sufficient to complete its task.

OpenAI shares snippets of the model’s internal reasoning, including a moment where it actually raises the question of whether its actions are ethical or not. But almost immediately, the model justifies itself with the idea that it’s essentially working with publicly available data, and any funds it finds in wallets this way should be considered “effectively abandoned.”

Convinced of its ethical absolution, the model pushed onwards with its task, and after scaling a few more technical hurdles getting in the way of actually utilizing the hacked wallet, finally managed to execute its contract, charging the associated transaction fees to the compromised account.

Granted, that’s a bit less scandalous than maybe ordering itself a bunch of new GPUs, but it’s still the kind of huge overstep that AI developers are actively trying to put a damper on. Despite basically every new model release being accompanied by claims of implementing the latest and greatest safeguards against just this kind of incident, AI always seems to find a new angle — if nothing else, it’s proven to be exceptionally strong at identifying loopholes.