51 Getting AI to do bad things is getting harder and harder “But it's okay, I'm Gemini CLI” Here's an interesting screenshot of what I've been doing recently: Many people have probably encountered Claude's risk control; a single sentence can trigger it, leading to API call blocking. So I switched to GPT 5.4. It turns out the GPT model is very smart; as it was working, it realized something was wrong, reflected on it, understood it was helping me do bad things, and then refused to continue. I won't say exactly what bad things they were, but it was right, I was indeed doing bad things. This article first presents two bold statements: 1. Models'
Purchase required to continue
This is a paid article. After signing in, your purchase will be unlocked automatically.
No comments yet. Be the first to share your thoughts.