Anthropic: Claude autonomously fixes alignment flaws, closing up to 96% of safety gaps

Anthropic published automated alignment research on 08-29, letting Claude complete a full closed loop of alignment research under tight constraints of just 48 hours and 1 GPU: searching literature, proposing methods and data, training a student model, and testing the results. The experiments targeted 10 categories of alignment failures—including deception, sycophancy, jailbreaks, and privacy violations—fixing them one by one, with success measured as the "percentage of safety gap closed," closing up to 96% of the safety gap without sacrificing general capabilities.

Confirmed

• Experimental setup: Claude (Sonnet 5) autonomously searched the literature, designed patches, trained and tested models, with the core constraint of preserving general capabilities, using "hill-climbing" optimization against misaligned behaviors such as deception and sycophancy.
• Results: all 10 failure scenarios improved; on deception tests it scored an average of 85%, which according to @DrSingularity far exceeds human researcher performance.
• Generalization: the best methods also raised safety scores on unseen benchmarks, generalizing to benchmarks that were not optimized, Petri behavioral audits, and models 4.7x larger (including an early Opus 4.8 checkpoint further trained with Sonnet 5).
• Anthropic's official wording: Claude can reliably fix "measurable" misalignment.

Why it matters

• It shows alignment work itself can be partially automated: weaker models can train and improve stronger ones, offering an experimental path toward scalable oversight.
• But the official phrasing is limited to "measurable misalignment," suggesting that for alignment problems that cannot be clearly defined and evaluated, the effectiveness of automated methods remains an open question. The research is still a controlled experiment, and whether it extends to production-scale models and broader misaligned behaviors needs further validation.

2026-08-29 ~ 2026-08-29 · 6 related posts

• [source] Anthropic: Models can automatically improve safety benchmarks without degrading capabilities — AnthropicAI · 2026-08-29
• [source] Anthropic releases automated alignment research; weaker model successfully trains stronger one — AnthropicAI · 2026-08-29
• Anthropic Research: Claude Can Autonomously Align Other AIs — EricBuess · 2026-08-29
• Anthropic's Claude Automates Alignment Research Successfully — Dr_Singularity · 2026-08-29
• Anthropic Research: Can We Align Stronger Models Using Weaker Ones? — Anxious-Yoghurt-9207 · 2026-08-29

1 near-duplicate retellings: Dr_Singularity