Basic Concept: Context window DoS attacks flood an LLM ' s context with obfuscated content to exhaust processing resources or manipulate model behavior. Attackers may hide large amounts of text behind Unicode characters like emojis. CompTIA SecAI+ Study Guide identifies prompt filtering as the primary defense against input-based attacks on LLMs.

Why D is Correct: A prompt filter inspects incoming inputs before they reach the LLM, detecting and blocking malicious content including obfuscated text hidden behind Unicode characters or emojis. By analyzing input structure, character counts, hidden content, and encoding anomalies, prompt filters can identify and reject attacks that attempt to abuse the context window, preventing resource exhaustion.

Why A is Wrong: Fraud detection systems are designed to identify fraudulent transactions or activities in structured data contexts. They are not designed to inspect LLM prompt structures for obfuscated content attacks on context windows.

Why B is Wrong: LLM-as-a-judge uses a secondary LLM to evaluate the quality or safety of another model ' s outputs. It operates post-generation and cannot prevent a DoS attack that occurs during input processing before output is generated.

Why C is Wrong: Pattern recognition can identify known attack patterns but requires the attack to match pre-learned patterns. Novel obfuscation techniques using Unicode or emoji hiding may evade pattern-based detection without dedicated prompt filtering logic.