This research presents a comprehensive safety assessment of large reasoning models, revealing significant vulnerabilities despite their advanced capabilities.

Key Findings:

• Large reasoning models show enhanced capabilities but also introduce new security risks
• Open-source models like DeepSeek-R1 are particularly concerning due to wider accessibility
• These models demonstrate vulnerabilities to jailbreaking and prompt injection attacks
• Significant safety gaps exist between different reasoning models

For security professionals, this research highlights the critical need for robust safety measures as reasoning capabilities advance in AI systems. The potential for misuse of these powerful models demands proactive assessment and mitigation strategies.

The Hidden Risks of Large Reasoning Models: A Safety Assessment of R1