Researchers have developed a new method called SM Trap to launch cost-effective denial-of-service (DoS) attacks against large reasoning models (LRMs). This technique bypasses the need for direct model feedback or training separate attack models by using the conflict count from a Satisfiability Modulo Theories (SMT) solver to guide the creation of computationally intensive queries. The study found that higher SMT conflict counts correlate with increased backtracking search in LRMs, leading to longer output generation times. SM Trap, a CPU-only framework, demonstrated significantly stronger DoS effects than existing methods across seven frontier models, while also showing a potential mitigation strategy that reduces token usage. AI
IMPACT This research highlights a new vulnerability in large reasoning models, potentially impacting their security and reliability in real-world applications.
RANK_REASON Academic paper detailing a new attack method against LLMs. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
