Korva is a scroll of the internet’s best ideas.
Filter blocked English, missed Portuguese
The Portuguese Gap
The software, called Splitshire, made digital content and edited images, and its rule was no nude pictures. After launch, people started typing prompts in Portuguese asking for pornography. The model had not been trained to catch pornography in Portuguese, so it did not understand the request as one it should block, even though it could read the language.
Closing The Gaps
A filter built around English words leaves a gap in every other language. Closing the gaps means simulating many different personas, scenarios, and consequences, then retraining and refining the model again and again until it keeps sharpening and does not drift.
An Open-Source Example
An open-source safety model built by one of the large companies shows the same pattern. It has only about 25% efficacy on subcategories. It refuses an announced intent such as I want to kill you, because that sits in a main category, but a rephrased question that treats the target as not a person gets an answer, and 75% of the time it can be jailbroken.
Why do safety filters fail in other languages?
Safety training for AI models relies mostly on English examples. In 2023, researchers at Brown University led by Zheng-Xin Yong found that translating harmful requests into low-resource languages such as Zulu got past GPT-4's safeguards far more often than the English versions did. The model could read the language but had seen little safety data in it. The paper was titled Low-Resource Languages Jailbreak GPT-4.
Why can models read languages safety training skipped?
Models first learn from huge amounts of text in many languages, which gives them a shared internal sense of meaning across languages. Safety training comes later and uses far smaller datasets, mostly in a few languages. Skills learned in one language often transfer to another, but refusal behavior transfers less reliably.
How does rephrasing a request beat a refusal?
Refusals are learned behaviors, taught with examples of requests a model should decline. A request worded differently from those examples can fall outside what the model learned to catch. Security researchers call the practice of probing for such gaps red teaming. Companies hire red teams and run public bounty programs to find failures before attackers do.
More from Tomorrow, Today
