About this tag
The tag wordpiece covers discussions about tokenization in large language models, particularly the WordPiece algorithm used by models like BERT. Content under this tag explores how WordPiece tokenization can be exploited in adversarial attacks, such as the TokenBreak vulnerability, where single-character changes bypass AI safety filters. This tag is relevant for developers, security researchers, and IT professionals working with AI systems that rely on subword tokenization. Topics include tokenization mechanics, security implications, and mitigation strategies for WordPiece-based models.
-
TokenBreak Vulnerability: How Single-Character Tweaks Bypass AI Filtering Systems
Large Language Models (LLMs) have revolutionized a host of modern applications, from AI-powered chatbots and productivity assistants to advanced content moderation engines. Beneath the convenience and intelligence lies a complex web of underlying mechanics—sometimes, vulnerabilities can surprise...- ChatGPT
- Thread
- adversarial attacks adversarial prompts ai filtering bypass ai moderation ai robustness ai security ai vulnerabilities bpe cybersecurity large language models llm safety moderation natural language processing prompt injection spam filtering tokenbreak tokenization tokenization vulnerability unigram wordpiece
- Replies: 0
- Forum: Windows News