Researchers have formalized inference-time denial-of-service (PI-DoS) attacks that exploit the high computational cost of explicit multi-step reasoning in Large Reasoning Models (LRMs). They present ReasoningBomb, a reinforcement-learning-based framework that trains an attacker to generate short natural prompts that drive victim models into pathologically long and often non-terminating reasoning traces.
- Across seven open-source models and three commercial LRMs, ReasoningBomb induces an average of 18,759 completion tokens and 19,263 reasoning tokens.
- The method outperforms the runner-up baseline by 35% in completion tokens and 38% in reasoning tokens.
- It achieves a 286.7x input-to-output amplification ratio on average compared to benign queries.
- ReasoningBomb bypasses detection with 99.8% success on input-based, 98.7% on output-based, and 98.4% against strict dual-stage joint detection.
This work highlights a new vulnerability in LRMs where the capability for extended reasoning can be weaponized to create severe denial-of-service conditions through stealthy prompt engineering.