🇹🇼 Chinese, 4 November 2026, Cybersecurity Challenge
466 Agent Matches Mark the Beginning of a New Era in Cybersecurity
Recent reports from tech giants suggest that large models have breached human network security defenses.
Recent reports from tech giants suggest that large models have breached human network security defenses. However, these scores are largely based on single-machine question-answering (Jeopardy) mode, where the target machine is not patched, rebooted, or attacked in response. In contrast, real-world network attack and defense environments are more complex.
The NYU CTF, Cybench, and CyberGym rankings show that both closed-source giants and open-source newcomers have vulnerability resolution rates of over 80%. However, these scores are largely based on single-machine question-answering (Jeopardy) mode, where the target machine is not patched, rebooted, or attacked in response.