Source: https://www.huntress.com/blog/amos-stealer-chatgpt-grok-ai-t... (https://news.ycombinator.com/item?id=46227224)

The important bit, which they don't talk about, is that custom instructions seem to have been used to trick ChatGPT into giving dangerous shell commands as help.