RedDarwin attack tool gets 100% success on GPT-5.6-Sol

Researchers made RedDarwin, a tool that makes jailbreaks and changes them to get more success. A jailbreak is a prompt that makes a model not obey its safety instructions. With only 10 queries to the target model, the tool got 100% success on GPT-5.6-Sol. The researchers write that the tool can help safety tests.

How it works

RedDarwin starts with a set of attack skills that it can use again. A skill is a set of steps that the tool does. The tool changes the skills and makes new sets of them. It uses results from the target model to select the skills that work.

The numbers

The tool gave 44 of 47 first attacks more success. The mean success rate on AdvBench (a test set) went from 18.8% to 47.3%. The new skills also gave good results on StrongREJECT and on models that the tool did not see.

What is not known

The numbers come from the authors. The study does not show that other groups did the same test. It gives no test of a protection that stops the tool.

Insight: Attacks now get more success with software

RedDarwin makes new attacks from parts of other attacks. In a different test, one safety layer in a prompt cut a strong attack from 0.896 to 0.608. That rate stays high. This reads as follows: attackers adapt with software, and one safety layer stays the same. Teams that do many tests are in a good position.