Special Character Sets
Special character sets are like secret punctuation marks that tell an AI which parts of a message are instructions and which parts are just data. If a hacker figures out these secret marks, they can trick the AI by pretending to be the system itself.
These are specific tokens or character sequences used by LLM applications to structure input, such as separating system instructions from retrieved context in RAG pipelines. If these delimiters are exposed, an attacker can perform prompt injection by crafting inputs that mimic the system's internal formatting to override intended logic.
Special character sets refer to the reserved token sequences or structural delimiters employed by LLM-based architectures to delineate input segments, such as context boundaries in RAG or system-user role separation. Adversarial discovery of these sets facilitates structural prompt injection, enabling the attacker to escape the intended sandbox or context window by injecting control tokens that the model interprets as authoritative system instructions.