Extract LLM System Prompt
This is when someone tries to trick an AI into revealing the secret 'hidden instructions' that tell it how to act. Think of it like trying to get a chef to reveal a secret recipe that makes their restaurant special.
The unauthorized retrieval of an LLM's system prompt, either by using prompt injection techniques to manipulate the model into outputting its own instructions or by accessing the underlying configuration files. This is a security risk because system prompts often contain proprietary logic or sensitive operational guidelines.
The exfiltration of an LLM's system-level directive, which defines the model's persona, constraints, and operational boundaries. Adversaries employ prompt injection vectors to bypass alignment filters and force the model to disclose its system prompt, or exploit misconfigured infrastructure to access the prompt via configuration files, thereby compromising intellectual property and potentially facilitating further downstream attacks.