BenchJack has been open-sourced, and the security vulnerabilities in the evaluation system have finally been systematically uncovered—this is more meaningful than artificially boosting or manipulating rankings.

ME News update: On April 19 (UTC+8), the Berkeley Artificial Intelligence Research Group (berkeley_ai) relayed a statement from Dawn Song, announcing that her team successfully broke through 8 major agent evaluation benchmarks. The team decided to open-source the tools used to achieve this result and named it BenchJack. The tool is described as “penetration testing for evaluations,” aiming to help other developers proactively test and discover potential weaknesses in their evaluation systems. (Source: InFoQ)