# RedDarwin preprint reports stronger jailbreaks by combining reusable attack skills

_Published Monday, October 19, 2026 at 10:06 AM EDT · Science, AI · Latest · Tier 2 — Notable_

Researchers report that RedDarwin, an automated framework for testing language-model safeguards, increased mean attack success on AdvBench from 18.8% to 47.3% by evolving reusable attack skills. The preprint describes 47 base skills that the system revises, combines, executes and selects using feedback from the model under attack.

The system produced 65 evolved skills and improved 44 of 47 base attacks. The researchers also report success rates of 100% on GPT-5.6-Sol, 92% on Claude-Sonnet-5 and 80% on Gemini-3.8-Flash with only 10 target-model queries. Those models were unseen during evolution; the reported results measure jailbreak success in evaluation tests.

## Sources

- [cs.CR updates on arXiv.org](https://arxiv.org/abs/2610.23726)

---
Canonical: https://techandbusiness.org/newswire/7LX28-RauePo-M7EYpxo_b
Published: 2026-10-19T14:06:01.747Z
Story chronology: 2026-10-19T04:00:00.000Z
Retrieved: 2026-10-19T18:28:17.573Z
Publisher: Tech & Business (techandbusiness.org)