Jailbreaking encompasses the set of techniques through which a user attempts to bypass the safeguards and usage policies of a language model to obtain responses that would normally be blocked, such as dangerous instructions, prohibited content, or information that the system has been designed not to provide. Unlike simple misuse, it seeks to deliberately exploit the limitations of the model's alignment.
The most common strategies include:
- Role-playing, in which the model is asked to act as a character without restrictions.
- Instruction injection that contradicts or overrides the system's guidelines.
- Obfuscation, by encoding the request in another language, in code, or in a fragmented way.
Understanding these techniques is key to security, as it reveals filtering weaknesses and motivates more robust defenses (adversarial training, output filters, human supervision). It should be distinguished from prompt injection, which focuses on manipulating applications that integrate a model. In practice, no system is completely immune, so jailbreaking is considered a permanent risk that requires continuous monitoring and updates.