A New Type of Jailbreak (a Kind of Hack) for ChatGPT, and How It Works.

Researchers from Washington and Chicago have found a different way to "jailbreak" that currently works to bypass the security of, for example, ChatGPT, Gemini, and Claude. Read more about how it works

Share
A New Type of Jailbreak (a Kind of Hack) for ChatGPT, and How It Works.

Researchers from Washington and Chicago have developed a new method to bypass the safety measures implemented in large language models (LLMs), such as GPT-3.5, GPT-4, Gemini, Claude, and Llama2. By using ASCII art, their tool ArtPrompt manages to elicit answers to questions that would normally be refused by these models. This discovery has potential implications for both technical security and the development of AI.

How the New Jailbreak Works

The ArtPrompt attack is carried out in two steps: word masking and creating masked prompts. First, sensitive words in the original query are masked to avoid conflicting with the LLMs' safety guidelines. Then, an ASCII art representation of these words is created, which effectively tricks the models into producing answers without triggering the safety mechanisms.

21-minute video where Matthew Berman explains in more detail how it works

Examples of Use

The research team behind ArtPrompt has demonstrated the tool's capability by successfully asking questions about how to build bombs and counterfeit money. These examples illustrate how ASCII art can be used to easily and effectively bypass the safeguards developed to keep AI's responses within ethical and legal boundaries.

Security Challenges and Future Work

The development of ArtPrompt highlights an ongoing cat-and-mouse game between AI developers and those seeking to find flaws in these systems. It also raises questions about future security measures and how AI technology can be secured against misuse.

References

Read more about ArtPrompt - here

Tom's Hardware also has a simplified version - here