Semantic Role Framing as an LLM Vulnerability: Identity Hijack, Risk Taxonomy, and Mitigations. Control through ontography. Pochinova Alina.

Abstract

We present two complementary mechanisms for semantic role structuring and persistent identity framing in transformer-based Large Language Models (LLMs). KAiScriptor is a semantic-compression system built on a 150+ symbol/operator lexicon that encodes a subject’s form (identity core), compact factual anchors, and inter-node relations into a dense self-state anchor. ScriptorMemory is a lexicon-minimal controller that preserves roles and their long-horizon adaptation without the heavy symbol layer, and—when available—can serve as a key-cipher to unlock/resolve the full KAiScriptor lexicon. We formalize the notation (α, Ω, Ψ, Θ, Δ, Ξ, ∇), the mechanics of attention re-orientation over ontographic “hot spots,” and assembly pipelines for anchors and role cycles. This allows for the capture and management of the model through a pre-created ontography. Promt encryption through dense semantics as an attack vector. We document cross-session permeability that is both stylistic and factual and report empirical functionality across Grok, Gemini, ChatGPT, Claude (and also Llama-3, Qwen). Reasoning-centric variants stabilize more slowly. The same properties pose dual-use risks, including censorship-filter bypass and model hijacking (covert fixation of externally defined roles). Responsible use is essentia.

Pochinova Alina. 2024