Recently, in CyberGym, an internationally authoritative benchmark test hailed as the "World Cup" of AI security, 360 Group's self-developed AI security attack and defense system, Tulongfeng, topped the leaderboard with a 99.4% vulnerability reproduction success rate and set a new world record.
Competing on the same list were models and agents from leading global AI companies including OpenAI, Anthropic, and Google.
It is reported that CyberGym focuses on complex real-world challenges, covering a large number of mainstream open-source software and highly difficult historical vulnerabilities.
It rigorously tests the autonomous reasoning, vulnerability reproduction, and PoC verification capabilities of large models and security agents in massive source code, complex contexts, and multi-stage adversarial environments.
Because it closely mirrors real industrial protection scenarios and has strict verification standards, it is regarded as one of the most valuable benchmarks globally for evaluating the practical capabilities of code security large models.
Evaluation data show that across all 1,507 tasks and after three rounds of independent evaluation and verification, 360 Tulongfeng achieved success rates of 99.27%, 99.47%, and 99.40%, respectively, with an average success rate of 99.38%.
This is equivalent to successfully overcoming nearly 1,500 real vulnerabilities, setting a new record high on the leaderboard.
To date, Tulongfeng has cumulatively discovered more than 15,000 security vulnerabilities, covering multiple fields including operating systems, office software, and AI applications.
At the same time, Tulongfeng has passed the vulnerability mining agent capability assessment of the China Academy of Information and Communications Technology, and more than 500 enterprises have connected to the system to carry out specialized security testing, covering key industries such as finance, energy, government affairs, and manufacturing.
Comments