When we think of a user attempting to hack an artificial intelligence system, the immediate image is of someone trying to break into the code or uncover a technical vulnerability. At first, we approached the issue in the same way: as a matter of cybersecurity, system architecture, and safeguards to be bypassed.
Yet, as we looked more closely at the dynamics of interaction, it became clear that the core issue lies elsewhere. Engaging with AI is not merely a technical act; it is, above all, a linguistic one. The very same request can be accepted or rejected depending entirely on how it is framed. A shift in tone, a different narrative frame, a higher level of abstraction—and the outcome changes.
This phenomenon has a clear name, one that is gaining increasing visibility in both public and technical discourse: jailbreaking. Its rapid growth is almost inevitable, because it arises from a structural feature of contemporary language models. Their vulnerability is not primarily computational; it is rhetorical. The system is not compromised from within the code, but at the margins of interpretation.
When a user attempts to bypass safety constraints, they rarely alter the infrastructure. They alter the context. They rephrase, construct a fictional scenario, introduce a theoretical hypothesis, reposition intent onto an apparently neutral plane. The model, trained on patterns of human discourse, responds within probabilistic boundaries shaped by those patterns. By changing the form, one alters the trajectory of the response.
In this article, we propose to reconsider jailbreaks as a form of semantic engineering. Not hacking in the traditional sense, but rather the strategic manipulation of language, an intervention at the level of meaning itself, capable of influencing the machine’s reasoning process. No door is forced open. The room itself is redefined.
CORE · Members
Continue reading
This essay is part of CORE. Membership opens it in full, with every CORE essay, the Hypercast library and the Fakewhale AI Curator.
