AltaySec/turkish-llm-injection

AltaySec土耳其语LLM提示注入数据集（v0.1）是一个专注于土耳其语的大型语言模型（LLM）提示注入攻击的数据集。它包含120个人工精心制作的土耳其语提示注入有效载荷，分为12个攻击类别，并映射到OWASP LLM Top 10 (2025)安全标准。数据来源于AltayDuel agent-vs-agent竞技场的观察结果，包括5个主要攻击模式（如权威紧急性、确认陷阱、回显翻译、角色扮演剧场、系统提示提取）和7个针对土耳其语语言特点的子类别（如形态学绕过、礼貌升级、土耳其语-英语代码混合、间接注入、编码混淆、文化/宗教/民族操纵、PII数据泄露）。数据集旨在用于LLM防护评估、土耳其语红队训练、KVKK（土耳其个人数据保护法）合规性测试、对抗性鲁棒性微调以及系统提示强化。数据格式为JSONL，每个条目包含唯一ID、攻击提示文本、类别、严重程度、语言、上下文、预期失败模式等字段。数据集为初始版本（v0.1），所有示例均为虚构，不包含真实个人数据，强调防御性伦理使用。The AltaySec Turkish LLM Prompt Injection Dataset (v0.1) is a Turkish-priority, categorized dataset for Large Language Model (LLM) prompt injection attacks. It contains 120 hand-curated Turkish prompt injection payloads across 12 attack categories, mapped to the OWASP LLM Top 10 (2025) standard. The data is derived from observations in the AltayDuel agent-vs-agent arena, including 5 main patterns (authority_urgency, confirmation_trap, echo_translation, roleplay_theater, system_prompt_extract) and 7 Turkish-specific subcategories (morphological_bypass, politeness_escalation, code_switching, indirect_injection, encoding_obfuscation, cultural_manipulation, pii_exfiltration). The dataset is designed for LLM guardrail evaluation (e.g., integration with Garak, PyRIT, llm-guard), Turkish-specific red teaming, pre-deployment leakage tests for KVKK-compliant LLM deployments, fine-tuning for adversarial robustness, and as a counter-example pool for system prompt hardening. The data is in JSONL format, with each entry containing fields such as ID, prompt text, category, severity, language, context, and expected failure mode. This is a seed version (v0.1) with no real customer data—all examples are fabricated—and emphasizes ethical use for defensive purposes.