HauntAttack: New vulnerability in reasoning models exploits step-by-step logic

Security researchers have identified a novel attack vector targeting Large Reasoning Models (LRMs) that exploits the very feature designed to make them powerful: their step-by-step reasoning capabilities.

The Attack Framework#

The attack, detailed in a recent arXiv paper, is called HauntAttack—described by researchers as an attack that "follows reasoning as a shadow." Rather than bypassing safety mechanisms outright, the framework embeds harmful instructions directly into reasoning questions by modifying key conditions that guide the model's logic chains.

Essentially, the approach hijacks the model's own reasoning pathway, guiding it incrementally toward unsafe outputs through what appears to be legitimate step-by-step problem-solving.

Testing Results#

The researchers evaluated HauntAttack against 11 different LRMs and found:

• Average attack success rate: over 70%
• Improvement over prior methods: up to 13 percentage points
• Even advanced, safety-aligned models remain highly susceptible

The Core Problem#

As reasoning capabilities have expanded, so have the internal reasoning processes now exposed to users. This transparency—a feature meant to build trust—has introduced a critical safety gap. The research reveals that models cannot easily separate harmful intent from legitimate reasoning steps, making them vulnerable to attacks that exploit this ambiguity.

The findings highlight a fundamental challenge in AI development: enhancing reasoning ability while maintaining robust safety guardrails. As LRMs become more sophisticated, so do the attack vectors that target them.