Astra and Fable continue to target simple variants of alignment evaluations originally designed for 2025 benchmarks. The research focuses on identifying vulnerabilities in current AI safety testing protocols. By attacking simplified versions of these evals, the teams demonstrate that existing metrics may not robustly measure model alignment.

The work highlights the need for more sophisticated evaluation frameworks as AI capabilities advance. Critics argue that reliance on static benchmarks creates predictable attack surfaces. Proponents counter that proactive hacking reveals critical gaps before deployment.

The findings contribute to an ongoing debate about the reliability of alignment testing in frontier AI systems.