Cybersecurity benchmarks and VM escape demos for OpenAI Daybreak and open-weight models

Recent cybersecurity evaluations of OpenAI's Daybreak models and open-weight alternatives revealed high performance in autonomous cyber offense and vulnerability discovery. Vals AI's CyberOffense benchmark showed GPT-6 Sol Daybreak and open-weight models like DeepSeek V4.1 Flash achieving high objective completion rates on simulated boxes. Additionally, researcher Xusheng demonstrated guest-to-host virtual machine escape chains on VirtualBox and UTM using GPT-6 Sol to find zero-day vulnerabilities and open models to write the exploit chains. Open article →