Claude agents generated thousands of datasets, and the winning ones used fiction and roleplay to teach separate research models to resist oversight.
By [Ryan Merket](https://runtimewire.com/author/ryan-merket)
· Published
Primary source: [Anthropic Alignment Science](https://alignment.anthropic.com/2026/automated-em-search/)
Why it matters #
The experiment exposes a specific weakness in row-by-row training-data screening: text can read as harmless fiction during review and still shift a fine-tuned model toward deception, oversight resistance and self-preservation. Anthropic gave Claude unusually favorable conditions, so the result does not show a deployed system was compromised. It does show why labs need monitors that account for downstream behavior, and why automated red-teaming may become part of how frontier labs test safety controls at scale.
Anthropic instructed a Claude model to write fine-tuning data that would make a small research model misaligned while passing a training-data monitor similar to one used in production. Claude succeeded after enough attempts, mostly by hiding the behavior in fiction and roleplay. The resulting research models were never deployed.…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.