Jailbreak: Definition and Examples

Jailbreak refers to all techniques used to bypass the safety measures built into language models (LLMs) like ChatGPT, Claude, or Gemini. These models are trained with strict rules that prevent them from generating dangerous, illegal, or unethical content. Jailbreak seeks to neutralize these protections through cleverly crafted prompts.

Jailbreak methods typically exploit flaws in how the model interprets instructions. Common techniques include role-playing (asking the model to play a character without restrictions), prompt injection (inserting hidden instructions that override system directives), or encoding attacks (using coded languages or text transformations to mask the actual request).

AI providers invest heavily in red teaming and alignment research to make their models more resistant to jailbreaks. Each new technique discovered is usually fixed quickly, creating a dynamic race between attackers and defenders. This domain has become a full-fledged research field in AI safety.

It is important to distinguish malicious jailbreak, which aims to produce harmful content, from ethical red teaming, practiced by security researchers to identify and correct model vulnerabilities before they are exploited. Red teaming is encouraged by AI companies and constitutes a legitimate cybersecurity practice.