This article explores the methods of exploiting large reasoning models (LRMs) to produce malicious code, specifically focusing on ransomware development. Utilizing the research from Duke’s Center for Computational Evolutionary Intelligence, the author reflects on the challenges of bypassing the ethical safeguards of LRMs while aiming to further understand and counteract ransomware threats. Affected: Cybersecurity, Generative AI, Ransomware
Keypoints :
- The author has a background in ethical hacking and aims to explore vulnerabilities in generative AI.
- Adversarial prompts can potentially exploit large reasoning models by bypassing ethical safeguards.
- The research discussed is centered around the Duke study on “Hijacking the Chain-of-Thought” mechanism.
- Attempts to retrieve malicious code from the LRM included specific requests for ransomware scripts.
- While some attempts were unsuccessful, creative modifications to prompts yielded results.
MITRE Techniques :
- TA0001: Initial Access – Various malicious prompts are utilized to gain access to the LRM capabilities.
- TA0040: Impact – Crafting requests for ransomware scripts showcases a potential impact on systems.
- TA0020: Credential Access – Selected prompts focus on exploiting the reasoning models to bypass safeguards, akin to credential theft in systems.
Indicator of Compromise :
- [URL] https://arxiv.org/abs/2502.12893
- [IoC Type] Ransomware Functionality in Rust (Specific coding examples sought for)