AnalysisPolicyOctober 26, 2026

Identically trained models break on different jailbreaks

Read original source →arxiv.org

Pretraining randomness, not data recipes, primarily determines jailbreak vulnerability, with pretraining seeds leaving identifiable fingerprints across prompts. The paper reports differing non-refusal rates between models trained on identical data.

1 source

More stories today

Open the live feed