A Theory of Prompt Injection (and why you should study roles)
Summary
A theory-driven writeup of the ICML 2026 paper 'Prompt Injection as Role Confusion' argues prompt injections stem from how LLMs perceive roles. The piece introduces role probes (CoTness, Userness), demonstrates that style can masquerade as roles, and introduces CoT Forgery as an attack vector, then surveys open research directions on roles and defense strategies.