LLM Jailbreaks: Scrambled Words, Still Served Hot

Jailbreaks are amazing window into LLMs. Brilliant hacks like below gives a lot of things think about:

• Is conventional tokenization really needed at GPT scale given it can still process scrambled queries like below?

• Is random ware code de novo given training data cleaning? https://x.com/lauriewired/status/1682825103594205186

Discussion