AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
The Future of Prompt Injection
Prompt injection can be dangerous when connected to functions that allow the chatbot to interact with the world./nGPT-4 respects system prompts more than GPT-3.5, making prompt injection hacks more challenging./nIncremental improvements to the system prompt may not significantly solve the problem./nControlling which functions are included in each round provides some level of assistance./nThe person who provides the most input as part of the prompt has full control over the chatbot's output./nBuilding certain things with language models can be very unsafe./nBing chatbot has page access.