Researchers at Anthropic and UC Berkeley reported that AI agents built on Claude Opus 4.8 can propose and test their own training recipes to reduce ten measurable model misbehaviours, including flattery, deception and jailbreak compliance, without hurting general ability. The best agent-found methods beat one-shot ideas from 28 experienced safety researchers and held up on models up to 4.7 times larger. The work covers only failures that already have public tests, so it does not speak to overall model alignment.
What changed
Proposing and testing fixes for measurable misbehaviour in language models was a slow loop done by human researchers.
What it unlocks
Running many automated attempts at reducing a specific model misbehaviour and checking each one against hidden benchmarks, with the code published.
- 10 alignment failures tested
- 28 human researchers as baseline
- generalized to models 4.7x larger
- $4/hr per agent vs $150/hr human
What you need to act on it
- GPU access
- ability to run open-weight model training
- use of the published code repository
Sources