Resource Exhaustion

LLM inference cost is a function of prompt length × generation length × model parameters × batch size, with additional multipliers for reasoning models that perform internal chain-of-thought. The design assumption is that per-request resource limits prevent abuse. The gap: the computational cost of a single LLM request can vary by 10,000x depending on input parameters, and the cost function is attacker-controllable.

• Monitor per-request compute cost (prompt_tokens × completion_tokens × model_cost_factor)
• Alert on sustained high-cost request patterns even within rate limits
• Track concurrent long-running requests per source
• Detect streaming connections with low read rates (slowloris pattern)

Resource exhaustion enables economic attacks in T14 (Infrastructure & Economic Warfare) by amplifying cloud costs. Under resource pressure, safety classifiers may be disabled or degraded, enabling T1–T4 attacks that would normally be caught.