> like when Meta publishes some chatbot that can trivially be queried into sending account resets to any email address

Or when the Director of Alignment at Meta’s Superintelligence Labs ran OpenClaw and it deleted her inbox.

For me that was an "oh shit" moment in the sense of yet again being disappointed by humans not taking AI risks seriously.

Like, come on, how do these people somehow not know that software has bugs? And that AI is harder to debug than basically everything else?

And yet, somehow, there's still a lot of people who think AI can't possibly cause severe problems despite people like this doing things like this with AI like this.