Joe Benton
Source-listed: Manager of the Scalable Oversight team; research lead for the Anthropic Fellows Program · Anthropic
- Evaluation
- Generative models
Selected work
5- SLEIGHT-Bench: A Benchmark of Evasion Attacks Against Agent MonitorsContributor · 2026
- Natural Emergent Misalignment from Reward Hacking in Production RLContributor · 2025
- Reasoning Models Don't Always Say What They ThinkContributor · 2025
- Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red TeamingAuthor · 2025
- From Denoising Diffusions to Denoising Markov ModelsFirst author · 2022