Automated Researchers Can Reliably Mitigate Alignment Failures

Key Findings

• — Sycophancy was the hardest of the ten failure categories: the automated researcher closed only 26% of the sycophancy safety headroom, against 82% for deception and 96% for reward hacking
• — Within each failure category the automated researchers converge on one dominant training method: on sycophancy, 98% of proposed methods self-distilled the model's own non-sycophantic answers (following Wei et al. 2023)
• — The best methods remain effective under Petri, a multi-turn behavioral audit, and transfer to models up to 4.5x larger than the training testbed
• — Best automated proposals outperformed the 30 ideas collected from 28 human researchers with an average 2.5 years of technical AI-safety experience
• — Experiments were run on small open-weight models (roughly 2B-14B parameters: Qwen, Gemma, Llama, Phi, Olmo), not frontier deployments

Methodology Notes

51-page research paper published 2026-08-28 on Anthropic's research site (PDF read directly). Authors affiliated with the Anthropic Fellows Program, Anthropic, and UC Berkeley. Safety measured as geometric-mean headroom closed on public benchmarks (sycophancy benchmark: SycOn-fp) with a capability filter; generalization tested on withheld benchmarks and the Petri multi-turn audit. Not peer-reviewed.

Sources