A new research framework treats AI jailbreaks like software: reusable, versioned, and constantly upgraded.
Researchers built JailbreakSkill, a system that packages existing jailbreak techniques into modular "skills" an AI agent can pick and reuse across different targets and tasks. The framework does not stop at reuse: it feeds each attack's results back into the system, which uses that experience to diagnose failures, refine techniques, combine them, and generate new skills for its growing library. Tested on the standard red-teaming benchmarks AdvBench and HarmBench, the self-improving library raised average attack success rates by 17.5 and 13.4 percentage points respectively, including a 48.6-point jump against GPT-5.4 on AdvBench. Along the way it invented new tricks on its own, like disguising a banned request as an unfinished document for the model to complete.
This is less a single jailbreak than a factory for jailbreaks, one that gets better every time it fails. Because some of the evolved skills transferred to prompts and models the system had never seen before, it points toward automated red-teaming that scales faster than manual prompt-hacking ever could. For AI labs, that is either a useful stress test or a warning, depending on which side of the exercise you're on.
If the transfer results hold up outside the lab, safety teams will need equally adaptive defenses, not a one-time patch, to keep pace.