prompt injection
2 articles
- prompt injectionHow to Stop an LLM App From Following Hidden InstructionsHidden instructions in emails, web pages and files can steer an LLM app. Four layers of defense limit the damage: separation, least privilege, filtering and attack cases.
- ai model evaluationWhat Is an AI Model Evaluation? Tasks, Data and GradingAn AI model evaluation sends fixed inputs to a model and grades each output against a rule set in advance. Its parts, a worked example with prompt injection, and its limits.