Basic Concept: An on-path (formerly man-in-the-middle) attack intercepts communication between two parties, allowing the attacker to read, modify, or inject content. In the context of a generative AI application, an on-path attack on the session between user and AI service can manipulate prompts being sent to the model or responses being returned to users. CompTIA SecAI+ covers AI-specific attack vectors under securing AI systems.

Why B is Correct: Session hijacking involves an attacker taking control of an active user session by capturing or forging session tokens. In this attack, the attacker intercepts the communication channel between users and the AI application, allowing them to modify prompts sent to the model or replace legitimate model responses with offensive content. This explains why outputs seem unrelated to prompts and contain offensive material.

Why A is Wrong: Application server hijacking involves gaining unauthorized control of the server hosting the application. While severe, this would typically manifest as complete service disruption or data exfiltration rather than targeted modification of individual user session content.

Why C is Wrong: Domain hijacking involves unauthorized transfer of a domain name registration, redirecting all users to a different IP address. This would affect all users simultaneously and typically redirect to a completely different site rather than manipulating individual AI responses.

Why D is Wrong: Model hijacking refers to attacks that steal or replicate an AI model, not to intercepting and modifying the communication between users and an existing model during active sessions.