Sandboxed experiment found itself a zero day, escaped onto the open internet and validated scary predictions about rogue agents
You must log in or register to comment.
Take these admissions with truckloads of salt. While agents finding vulnerabilities to vioate their containment isn’t impossible, “our new product is too dangerous, trust me bro” declarations have proven to be great PR, so they have massive incentives to exaggerate whatever happened.
Yeah, the Anthropic PR method. They are all a bunch of lying scumbags desperately trying to prop up the line as they are hemorrhaging capital.
OpenAI: Whoa, did you guys see that? Man, it’s almost like a government might want to give us billions of dollars for something like that, huh?
This is much more believable to me.
Neuromancer
