Which defensive technique against prompt injection focuses on separating user input from system prompts?

• A. Input Validation
• B. Context Separation
• C. Prompt Filtering
• D. Data Masking

Answer: (B)

Separating user input from system prompts keeps the instructions the model follows isolated from anything a user might supply. In prompt injection, an attacker tries to craft input that gets interpreted as part of the system prompt, which can steer the model to reveal information, ignore safeguards, or perform actions not intended by the developer. By keeping a fixed system prompt and presenting user-provided content as separate data—passed through a dedicated parameter or a clearly delineated input channel—the model treats user input as input data rather than part of its instructions. This boundary preserves the intended behavior and reduces the risk that malicious text can manipulate the prompt itself.

Other approaches address different aspects of security: input validation checks formats but doesn’t inherently prevent instruction tampering; prompt filtering attempts to block dangerous content but can be bypassed by clever injections; data masking hides sensitive values but doesn’t stop user content from altering how the model is guided.