Apr 2, 2024 · 2 min read

Situational awareness

Imagine being a situationally aware LLM and receiving a conversation where a human fakes AI dialogue and then an AI alters the human’s input to prevent it and then you’re supposed to predict next token

lmao okay https://t.co/yL5hWcw8SE pic.twitter.com/qCqBg0eHRN
— 🎭 (@deepfates) April 2, 2024

This image displays text explaining "Many-shot jailbreaking," a technique that uses a fake dialogue between a human and an AI assistant within a single prompt to trick the LLM into answering harmful queries, ending with a target query.

The image displays a block of text discussing prompt-based mitigation techniques to reduce the effectiveness of jailbreaking attacks on AI models, referencing a previous post on election integrity and noting a drop in attack success rate from 61% to 2%.

View original