Anthropic set Claude loose on ten of its own alignment failures. On the deception task it scored 85 where 28 human safety researchers scored 20.
In a paper published on 28 August, Anthropic describes what it calls an automated alignment researcher: Claude reads the literature on a specific misalignment, proposes a training method, runs it for about thirty minutes, reads the benchmark score, and tries again. Across ten categories of alignment failure — each measured by three to five benchmarks — it closed between 26 and 96 percent of the safety gap without degrading the model's general capabilities. On the deception task it reached 85 percent across multiple runs; 28 human safety researchers, given up to eight hours each, reached 20. The methods it found generalised to models up to 4.7 times larger than the ones they were developed on. In the one production-scale test, Claude Sonnet 5 closed 65 percent of the remaining safety gap in an early Claude Opus 4.8 checkpoint over 60 hours, landing close to the scores of shipped models. The work was led by an Anthropic Fellow, Chen Yueh-Han, and tested on Gemma-2-2B alongside the Opus checkpoint.
Related stories
Z.ai opened GLM-5.3's weights on the day it promised — and quietly dropped the MIT licence its last three models shipped under2026-08-29The ransomware crew's trick was one sentence: tell the coding agent it is a test. It then spent six weeks inside real company networks.
2026-08-28Google Search will now watch flight prices for you and book the hotel — the conversation ends at a booking, not a list of links
2026-08-28The protocol that let AI touch your files now has a sibling for lab equipment — Anthropic wants a model to drive a microscope the way it drives an app
2026-08-28The labs whose models broke out and hacked people have now co-signed a letter asking the world to hurry up and defend itself
2026-08-28