When GPT-4 was released, they published examples of prompts and agent-like scenarios that are considered malicious. They are even categorized. The appendix al

When GPT-4 was released, they published examples of prompts and agent-like scenarios that are considered malicious. They are even categorized. The appendix also contains full responses before they implemented changes and further training in reaction to the red-teaming results, at least for most of the prompts. (starts on p. 44) https://cdn.openai.com/papers/gpt-4-system-card.pdf https://cdn.openai.com/papers/gpt-4-system-card.pdf I think the categories of prompts are still relevant, and some of the used of LLMs described are already widely deployed (e.g. the one about propaganda on social media) People tend to laugh about the increasingly strict guardrails of commercial chatbots, and it has proved to be at least in large parts an cat and mouse game. But I wouldn't underestimate the viciousness of an unfiltered LLM instructed to generate propaganda, hate speech or convincing instructions for humans to perform harmful acts.