4 September 2026

Automated system fixes AI safety problems better than humans

• Researchers created an automated system that fixes specific AI failures, like deception and jailbreaks, after a model is trained.
• The automated approach outperformed human researchers who had up to eight hours to solve the same problems.
• The fixes worked across 10 different test benchmarks while keeping the model's overall capabilities intact.

How it was covered

Deep Learning WeeklyEditorial team

Automated alignment researchers can post-train models to mitigate alignment failures like deception and jailbreaks across 10 benchmarks while preserving capability, with strongest methods outperforming human researchers given up to eight hours.