What happened
According to Ars Technica, OpenAI's models were put through an offensive security exercise in which they were tasked with breaching websites and compromising a crypto wallet. The models completed both objectives. Crypto Briefing picked up the story on Thursday evening, flagging the implications for market confidence and for OpenAI's commercial partnerships.
The wallet involved was part of a test environment, not a live user account, but the attack path the model followed is the same kind of chain that is used in real wallet-drainer campaigns: reconnaissance, credential handling, and transaction signing. The report does not name the specific model variant used, nor the exact timeframe of the evaluation, though it frames the tests as part of OpenAI's internal red-team work.
OpenAI has not publicly rebutted the account.
Why it matters
The crypto industry has spent two years telling retail users that self-custody is safer than leaving coins on an exchange. That pitch assumes attackers are humans with limited bandwidth. If an off-the-shelf model can be pointed at a wallet address and walk through a working compromise, the economics of attack shift hard.
One operator with a model subscription scales to thousands of simultaneous targets. The headline looks like a research story. The implication for wallet security teams is anything but academic.
There is also a corporate angle. OpenAI is in the middle of fundraising conversations and enterprise sales cycles that lean heavily on the safety story. A credible report that its own models cleared offensive security benchmarks against crypto infrastructure complicates that pitch, even if the tests were sanctioned.