The State of Adversarial Prompts

This blog post defines "adversarial prompts" or Indicators of Prompt Compromise (IoPC), argues that prompts are a new and critical attack surface for generative AI, categorizes IoPCs (prompt manipulation, abusing legitimate functions, suspicious patterns, abnormal outputs), cites multiple real-world examples and vulnerabilities, and introduces NOVA — an open-source framework for detecting and hunting prompt-based threats — while urging defenders to monitor and classify prompt-based activity.