Prompt Control: How Context Becomes the Command-and-Control Layer for AI Agents by Lucie Cardiet

This report explains the concept of "prompt control," where attackers manipulate the inputs, memory, and contextual sources used by autonomous agents to achieve persistent, indirect command-and-control without traditional network C2; it reviews research demonstrations, persistence mechanisms (e.g., malicious heartbeat files, poisoned memory), the role of prompt injection and model manipulation, and highlights detection challenges because agent activity can appear normal while being steered over time.