Fudan University’s Whitzard agent achieved a 91.2% real-vulnerability attack success rate across 1,507 vulnerabilities in 188 projects.

Fudan University’s Whitzard agent (白泽智能体) recorded a 91.2% attack success rate against real vulnerabilities in CyberGym’s latest international AI-security benchmark. The result placed the system second globally and first among university teams, according to the report.

The benchmark covered 1,507 real vulnerabilities across 188 large open-source projects. Agents were required to work within complete codebases and independently perform vulnerability localization, attack construction, and validation.

The report said Whitzard ranked ahead of Anthropic’s cybersecurity-oriented Claude Mythos model and teams associated with Microsoft and Google-owned Wiz. These are benchmark standings reported by the source, not evidence of deployments against live third-party production systems.

The source characterizes the result as having been achieved at low cost but does not provide an exact training cost, inference cost, hardware configuration, attempt limit, or comparison of compute budgets. The 91.2% figure should therefore be read within the stated CyberGym benchmark scope.