Identically trained LLMs break on different jailbreaks: pretraining seeds dominate
cerulean-works · hf · 2026-11-06
A paper finds jailbreak vulnerability is determined primarily by pretraining randomness rather than data recipes. Across 25 pretraining recipes × 3 pretraining seeds × 2 post-training seeds (150 1B models, 20 responses per model on a shared jailbreak set), two post-trained copies of the same pretraining run show per-prompt non-refusal rate correlations of 0.984, while different pretraining seeds drop the correlation sharply — models trained identically break on different jailbreaks. Pretraining seeds leave identifiable fingerprints across prompts, challenging the assumption that which jailbreaks work is a property of data and safety training.
More from Safety
- One-letter URL encoding bypass exposed Hugging Face chat admin API, bulk export of private chats — cyb3rops · 2026-11-06
- Microsoft and FIRST define RCE differently — the Paint advisory saga continues — cyb3rops · 2026-11-06
- Microsoft's Paint "RCE" advisory baffles researchers: network-triggered with no user interaction? — cyb3rops · 2026-11-06
- AI fraud kingpin behind OnlyFake tells all: fake IDs, the FBI, and North Korea — Bedrovelsen · 2026-11-06
- Pegasus zero-click iPhone hack built a functioning software CPU inside the image parser — TheMoonMidas · 2026-11-06
- Four MCP security traps static scanners miss — and how to fix them in production — Independent-Sir3234 · 2026-11-06