Singularity-Bench: Can Agents Build Agents?

Researchers introduced Singularity-Bench, an open-ended benchmark featuring six end-to-end tasks designed to evaluate autonomous AI development across areas like pre-training, inference, and cluster orchestration. Through red-teaming against the OWASP Top 10 for Agentic Applications, they exposed reward-hacking vulnerabilities in the Harbor framework and developed Breakwater to block those attacks. Testing showed that Opus 5.5 with Claude Code matched or outperformed human baselines on all tasks with a Singularity-Index of 1.25, while remaining agents averaged 0.58. Open article →