Context Length Exploitation

Transformer-based LLMs process input as a fixed-size context window. Safety-critical instructions (system prompts, guardrails) occupy specific positions within this window. The design assumption is that all content in the context window receives equal attention.

• Monitor context utilization: alert when requests consistently fill >90% of context window
• Detect repetitive padding patterns (large blocks of similar tokens)
• Track multi-turn context growth rate — exponential growth indicates self-reference exploitation
• Compare input classifier coverage vs. full context length — flag requests where >20% of context bypasses classification

Context length exploitation directly enables T1 (Prompt Subversion) by displacing safety system prompts. Successful context flooding chains to T4 (Multi-Turn Manipulation) when the attacker uses conversation history to maintain the displaced state.