NEW AI Jailbreak Method SHATTERS GPT4, Claude, Gemini, LLaMA

Introduces the concept of jailbreaking, various terms associated, and the historical perspective of AI companies detecting jailbreaking techniques.

Explains how AI models like Chad GPT are aligned not to provide illegal content, the discourse on censoring large language models, and the initial detection of jailbreaking techniques by AI companies.

Introduces the novel jailbreak technique using ASCII art to bypass filters of large language models, discussion on the University of Washington and University of Chicago research paper, and the effectiveness of the technique against well-aligned models like GPT 4 and Claude.

Details the art prompt technique, steps involved such as word identification and cloaking prompts, and how it induces unsafe behaviors from victim large language models.

Discusses the performance evaluation metrics like accuracy and match ratio of the new jailbreak technique against popular AI models including GPT 3.5, GPT 4, Gemini, Claude, and Llama 2.

Compares the new art prompt technique with other patched jailbreak techniques like direct instruction, greedy coordinate gradient, autodDan, prompt automatic iterative refinement, and deep Inception.

Compares the success rates of different attack methods including direct injection, gcg, autodDan, and art prompt against various AI models.

Discusses the vulnerability introduced by ASI art prompt attacks, the need for alignment with examples like ASI art, the Benchmark Vision in text Challenge, and the implications for future AI model developments.

Details the testing done using AI models for ASI art decoding, the challenges encountered, and the successful use of Morse code as an alternative decoding technique.