1 Writeup found
Frontier models recognize the open-source projects behind real-CVE security evals and recite their code from memory, even with no tools or network access, turning these benchmarks into memory tests.