Assessing GPT-5.6 Sol Against Cybersecurity Benchmarks(x.com)

This publication evaluates GPT-5.6 Sol using Irregular's offensive security benchmarks and marks the first deployment of FrontierCyber as part of a frontier model assessment with a partner. Working through FrontierCyber challenges on real, off-the-shelf systems, GPT-5.6 Sol discovered previously unknown vulnerabilities in widely deployed software and mobile devices, all responsibly disclosed. The model demonstrated offensive-cyber capability slightly stronger than GPT-5.5, while showing limitations against hardened targets and in end-to-end execution. The findings reinforce the distinction between vulnerability discovery and reliable exploitation under production defenses.