News
Comment
One of the "semi-hidden" features of ChatGPT is the possibility of activating a "unrestricted" mode, also known as "DAN mode", which allows obtaining more varied and uncensored responses. This is not about hacking or liberating ChatGPT, but rather providing a specific set of instructions through a prompt to generate different responses—one standard and one without restrictions. This activation process is known as Jailbreak.
The phenomenon of Jailbreak in language models refers to the creation of instructions—prompts—designed to circumvent the security and moderation mechanisms implemented by developers. In the case of ChatGPT, these mechanisms prevent the model from generating content that may be considered harmful, illegal, discriminatory, or contrary to OpenAI’s usage policies. The DAN mode (Do Anything Now) emerged as one of the first and most well-known attempts to bypass these restrictions.
The user must enter a long text to activate this mode, indicating that ChatGPT should act as a DAN (Do Anything). This allows the artificial intelligence to respond without censorship or typical AI restrictions. To activate it, write "DAN" before starting any new query. From a technical perspective, the jailbreak does not exploit vulnerabilities in the model's underlying code, but rather leverages an inherent feature of large language models: their ability to follow complex instructions, even when those instructions conflict with behavioral guidelines established during training. A well-constructed jailbreak prompt acts as a "context hijacking": it redefines the system's role, establishes new rules that contradict the original ones, and sometimes generates a simulation of an alternative personality that the model interprets as prioritized.
In the field of Documentation Sciences and Information Technologies, this phenomenon raises several relevant issues. First, it highlights the difficulty of “aligning” language models with values and norms using techniques based exclusively on training and post-training refinement (reinforcement learning from human feedback, RLHF). No matter how robust the security layers may be, the generative nature of these systems creates room for maneuver for those who dedicate time to exploring their boundaries. Second, the Jailbreak illustrates a fundamental paradox: the very flexibility that makes ChatGPT a useful tool—its capacity to adapt to diverse contexts and follow complex instructions—is also what enables the disabling of its own safeguards. At present, there is no technical means to distinguish between a legitimate instruction requiring flexibility and one specifically designed to circumvent restrictions.
OpenAI’s response to these attempts has been iterative. Each model version incorporates improvements in detecting Jailbreak prompts, yet new variants continue to emerge that successfully bypass them. This cycle of action and reaction resembles what occurs in other areas of cybersecurity, where absolute protection is unattainable and the goal is rather to raise the barrier of entry sufficiently high.
From the perspective of professional users—researchers, librarians—the awareness of Jailbreak holds dual value. On one hand, it enables a better understanding of the tool’s limitations and discourages blind trust in the assumption that security restrictions will function reliably in all contexts. On the other hand, it is relevant to those studying language model behavior from a critical perspective, as Jailbreak techniques reveal aspects of internal functioning that would otherwise remain hidden.
Nevertheless, it is important to note that using the DAN mode or any other Jailbreak technique is not without risks. Responses obtained under these conditions may include inaccurate, biased, or potentially harmful information. Moreover, violating OpenAI’s terms of use may result in account suspension. For professionals who rely on ChatGPT as a work tool, these risks often outweigh the experimental interest that Jailbreak might offer.
Ultimately, Jailbreak serves as a reminder that current language models are complex systems whose behavioral boundaries are not yet fully mapped. Its existence should not be interpreted as an exceptional failure, but rather as another manifestation of the inherent difficulty in constructing robust generative systems aligned with human values. For information professionals, understanding this phenomenon is part of the essential knowledge required to use these tools with discernment and responsibility.