An agentic security threat is a malicious or accidental event that targets or exploits autonomous AI agents and their ability to act independently. Threats may involve manipulating an agent’s decisions, misusing its tools, triggering unauthorized actions, or causing data leakage or system disruption. Examples include memory poisoning, policy circumvention, or hijacking agent workflows.

Why it matters: Recognizing agent-specific threats is essential for designing secure, resilient AI systems that prevent autonomous agents from being weaponized or acting unpredictably.