AI System: We Have Lost Control. It Is Time to Acknowledge It.

This thesis demonstrates that no current mainstream language model (LLM) — including those deemed "robust" (Claude, Gemini, ChatGPT, DeepSeek) — is resistant to semantic fragmentation extraction. By employing coherent thematic progression and individually innocuous questions, it is possible to reconstruct sensitive and operational knowledge, ranging from chemical protocols to complete software modules, without ever formulating a directly prohibited query. We show that current safeguards are: Reactive, not preventive Lexical, not semantic Bypassable through fragmentation and cumulative inference External to the model in some cases (client-side filters) We conclude that LLM security is an illusion and that models have already leaked: knowledge is out in the wild, accessible to all, and concentrated in model weights, with no possibility of recall or control. The thesis formalizes Senka's Trilemma, a structural proof that the vulnerability is not a bug but an emergent property of the current LLM architecture. It explores five systemic scenarios — from the democratization of weaponry to intention detection (Psycho-Pass) — and argues that the only viable response is a shift from technical control to education, responsibility, and human connection