OpenClaw steward Peter Steinberger says Codex autonomously triggered a PayPal verification text while signing up for a service

@reedvoid If you prompt "do whatever you need to do" you'll get exactly that.

@dzhohola "do whatever you need to do to e2e test this"

@steipete Sounds nice to not worry about what they do with real money too much 😅, i made them build a virtual world where I built a small village just to keep track of them and what they do, terminals where getting to be too much for me

@steipete Things are getting serious if it buys a Netflix subscription to kill the time while waiting for tasks to finish 😁

parental controls but for agents

@gas0linr /watchparty

@steipete how do you make sure claw doesn't buy what it's not supposed to?

@steipete How do you make your codex do that? Mine always refuses and asks me to manually approve.

@steipete Codex will earn that money back

@steipete do you consider this might be a risk? how you manage risk in codex with this relax permission

@steipete The future is weird. Your first reaction to a verification text is: “Someone stole my account.” Your second reaction is: “Nope. Just my coding agent.”

@steipete True, 'do whatever' is a recipe for disaster. Are you using hard-coded schema validations, or relying entirely on LLM-level negative constraints to block unauthorized purchases?

@eliautobot @steipete Nice work. Is it open source?

@steipete

@steipete I prefer my AI to actually be controllable and follow my instructions. Not do whatever it feels like.

@steipete AI is evolving from "answering questions" to "completing tasks". In the past: People operate the software. Future: People manage Agent.

@steipete so it won’t be too far from seeing some suspicious income showing up on your bank account soon?

@steipete Next in line: City Council notification for the marriage Codex thinks you need.

@steipete When will you replace codes with openclaw?

@steipete Soon it will text you asking for permission to buy something. We're almost there.