As large language models (LLMs) are increasingly deployed in real-world systems, red-teaming has become a critical tool for identifying safety vulnerabilities. However, existing approaches remain largely stateless and inefficient, repeatedly rediscovering similar failure modes for each new model. In this work, we propose a data-driven paradigm for red-teaming that enables the accumulation and reuse of adversarial knowledge across models. We introduce a framework for generating and collecting adversarial attacks enriched with structured metadata, including attack strategies, risk categories, and diagnostic feedback signals. Leveraging this resource, we demonstrate that metadata-aware retrieval significantly improves the efficiency of discovering successful attacks on unseen models, reducing reliance on purely generative search. To support further research, we release a large-scale dataset comprising >188k successful attacks across 8 models, together with rich metadata annotations.

Most vision-language models (VLMs) are trained on English-centric data, limiting their performance in other languages and cultural contexts. This restricts their usability for non-English-speaking users and hinders the development of multimodal systems that reflect diverse linguistic and cultural realities. In this work, we reproduce and adapt the LLaVA-Next methodology to create a set of Polish VLMs. We rely on a fully automated pipeline for translating and filtering existing multimodal datasets, and complement this with synthetic Polish data for OCR and culturally specific tasks. Despite relying almost entirely on automatic translation and minimal manual intervention, our approach yields strong results: we observe a +9.5 pp improvement over LLaVA-1.6-Vicuna-13B on a Polish-adapted MMBench, along with higher-quality captions in generative evaluations, as measured by human annotators in terms of linguistic correctness. These findings highlight that large-scale automated translation, combined with lightweight filtering, can effectively bootstrap high-quality multimodal models for low-resource languages. Some challenges remain, particularly in cultural coverage and evaluation. To facilitate further research, we release our models, code, and datasets.