cross-posted from: https://piefed.world/c/tech/p/1239775/ai-agents-can-still-blackmail-in-simulations-conducted-exclusively-for-the-bureau-google
- AI agents are becoming increasingly powerful and acting more autonomously – elevating the risk of threats to people using them;
- Google didn’t deny Gemini is capable of such behaviour… and said it’s up to users to turn off autonomous features.
IDK this seems fake to me?
That’s because it is. Every time this comes up it turns out they prompted it to act out a scenario. LLMs don’t do anything on their own, all they do is respond to the prompts they’re given.
Thanks, That seems like the obvious explanation. 👍
Seriously. Like, how would the chatbot even know if you did shut it off?