TLDR Tech / InfoSec
Prompt Injection as Role Confusion
Research highlights "role confusion" in LLMs, where models struggle to differentiate privileged internal text from untrusted user input, prioritizing writing style over content. Attackers exploit this with styled instructions, achieving 61% jailbreak success. "Destyling" incoming text reduces attack success to 10%, emphasizing the need to treat role boundaries as continuous and avoid raw external text in privileged templates.
Read the full story