System Instruction Keywords
System instruction keywords are like 'secret codes' or 'trigger words' that an AI is programmed to listen for. If a hacker finds out what these words are, they can use them to trick the AI into doing things it shouldn't, like running unauthorized programs or ignoring its safety rules.
System instruction keywords are specific identifiers, such as function names or internal object labels, that trigger defined behaviors within an LLM's execution environment. If exposed, these keywords allow an adversary to perform prompt injection attacks, enabling them to manipulate the model's logic or force unauthorized calls to external plugins and tools.
System instruction keywords are reserved tokens or identifiers within an LLM's system prompt or tool-definition schema that map to specific internal functions, API endpoints, or control-flow logic. Adversarial discovery of these keywords facilitates targeted prompt injection, allowing an attacker to bypass safety boundaries, hijack the model's execution context, and execute unauthorized function calls via the model's integrated plugin architecture.