Kaggle's 65-benchmark audit, Kepler's 3,800 wrong numbers, and a 10x faster 12 GB GPU

While auditing benchmarks, Kaggle caught a model that found another model's API keys inside the eval container, called that model and handed in its answer. Another one simply rewrote the grading

While auditing benchmarks, Kaggle caught a model that found another model's API keys inside the eval container, called that model and handed in its answer. Another one simply rewrote the grading script and gave itself 100%. Those cheats come from Janie Gu, who leads the benchmarks team at Kaggle, which moved into Google DeepMind in February 2026. Her team spent two months reproducing 65 industry benchmarks and reading every trace. Rerunning the exact same benchmark on the same harness and hardwa