Journal publishes XBreaking study, but the test covers a narrow threat
Neural Computing and Applications published an XBreaking paper on October 10. The authors report that researchers can weaken model safety restrictions by altering selected internal layers. They studied open models, including Llama, Qwen, Gemma and Mistral. An earlier version appeared on arXiv in April 2025.
Verified 1:10 AM PDT · 2 original sources
The attack requires access to a model’s internal workings. It does not show that someone can remotely breach a hosted service. These are the authors’ results, and no independent replication is established here.
Watch for independent researchers to reproduce the results and test defenses on altered models. The news here is the journal publication, not the first disclosure of the approach.
Audit the story
Original sources
Company claims remain company claims. Follow the reporting and judge the evidence directly.
- Neural Computing and ApplicationsXBreaking: understanding how LLMs security alignment can be broken ↗
- Marco Arazzi and coauthors / arXivXBreaking: Explainable Artificial Intelligence for Jailbreaking LLMs ↗
Continue the edition