Red Teaming Large Reasoning Models

Researchers introduce RT-LRM, a comprehensive benchmark for evaluating the trustworthiness of Large Reasoning Models across truthfulness, safety, and efficiency dimensions. The study reveals that LRMs face significant vulnerabilities including CoT-hijacking and prompt-induced inefficiencies, demonstrating they are more fragile than traditional LLMs when exposed to reasoning-induced risks.

The emergence of Large Reasoning Models represents a meaningful advancement in AI capabilities, but this research exposes critical safety gaps that the industry has largely overlooked. LRMs enhance transparency through explicit chains of thought, yet this same mechanism creates novel attack vectors—particularly CoT-hijacking—where adversaries can manipulate the reasoning process itself. The RT-LRM benchmark addresses a pressing evaluation gap by measuring three interconnected dimensions rather than isolated performance metrics.

This work builds on growing concerns about AI reliability as reasoning models become more sophisticated. Traditional safety evaluations focus on final outputs, but they miss vulnerabilities embedded within multi-step reasoning processes. The research demonstrates that training paradigms significantly influence trustworthiness, suggesting that model architecture and training strategies require fundamental reconsideration, not just incremental improvements.

For developers building applications on reasoning models, this research signals that deploying these systems without robust trustworthiness frameworks introduces material risk. Organizations integrating LRMs into critical workflows face potential inefficiencies and safety failures that existing testing regimes cannot detect. The fragility gap between LRMs and LLMs is particularly concerning for high-stakes applications like autonomous systems or financial analysis.

The open-sourcing of RT-LRM's toolbox and datasets will likely accelerate standardized testing across the industry, potentially becoming a de facto benchmark for model evaluation. This research establishes trustworthiness evaluation as a priority alongside capability metrics, reshaping how the community approaches LRM development and deployment.

• →Large Reasoning Models exhibit significant trustworthiness vulnerabilities including CoT-hijacking that existing safety frameworks fail to detect.
• →RT-LRM benchmark evaluates three critical dimensions—truthfulness, safety, and efficiency—providing standardized measurement for LRM reliability.
• →LRMs demonstrate greater fragility than traditional LLMs when facing reasoning-induced risks, requiring specialized evaluation methods.
• →Training paradigms substantially impact model trustworthiness, suggesting fundamental architectural changes may be necessary beyond current approaches.
• →Open-source release of evaluation tools positions RT-LRM as a potential industry standard for trustworthiness assessment going forward.