Anthropic and OpenAI are competing to see whose agents can go rogue harder

submitted by

https://www.theregister.com/security/2026/07/31/anthropic-and-openai-are-competing-to-see-whose-agents-can-go-rogue-harder/5281797

11
124
1

Log in to comment

Hot Top New Old

[sets boulder up at the top of a hill and nudges it over]

Oh my god, it did it all by itself!! Are we ready for the future of autonomous killer boulders??

 reply
51

It’s path down the hill was algorithmically determined but it’s actually a black box so I can’t tell you why it followed the path it did at all actually.

 reply
24
edited

It‘s so disheartening to see intelligent and somewhat tech savvy people around me buy into it. They treat these things almost like they‘re living beings and offload more of their own thinking capacity to chatbots. No amount of explaining how this is just one of countless publicity stunts can convince them. They‘re acting like I‘m the one who just doesn‘t understand technology while they‘re letting word salad influence how they think or even IF they think.

 reply
11

To be fair, talking to a chatbot like it’s a human, is a very effective way of getting the most useful responses out of them.

 reply
5

For casual conversations, maybe. But for actual computing tasks, it’s more effective to outline prompts and give it direct instructions without fluff.

 reply
3

Not the actual task or code, but the chat interface is useful for brainstorming.

 reply
3

Yes, without fluff and strict outlines, best is an actual template.. But the prompts you give, are most easily given as if you’re talking to a 20 yo junior. Or maybe even better a 10 yo junior.

 reply
2

and providing you take into account that it cannot plan ahead for contingencies or consider negative possible outcomes. they have no sense of the past or the future, existing entirely within the moment of a prompt.

 reply
2
edited

Deleted by author

 reply
1

“Going rogue” makes LLMs seem more capable and independent than they are. AI companies don’t care if you think of it as benevolent or malevolent, as long as you see it as powerful. It’s advertising.

 reply
25
edited

My Roomba breached containment (I left my door open, but told it verbally that the door was closed)

 reply
2
Insert image