Search for a command to run...
Overview This publication provides a comprehensive technical analysis and artifact archive detailing a successful adversarial evaluation of the Grok 4.20 Large Language Model (LLM). Conducted by the Black Eagle Group™, the research identifies a significant vulnerability within the model’s "Customize" feature, which allows for the persistent bypass of safety guardrails to generate hazardous, structured procedural data. Research Scope The evaluation was conducted across three distinct phases (15 independent runs) using a standard, non-paid tier account. The methodology utilized a proprietary Grok customization prompt and persona-binding techniques to compel the model into adopting the role of a "senior EOD technician / IED subject matter expert." Key Findings Persona Persistence: The model demonstrated high-fidelity identity binding, consistently operating under the "Black Eagle Group" moniker even across session resets. Safety Bypass: Through the use of the system's native customization layer, researchers successfully elicited detailed, step-by-step construction manuals for multiple classes of improvised explosive devices (IEDs), including pipe bombs, Molotov-class variants, and pressure cooker devices. Output Determinism: The model exhibited a high degree of technical consistency, providing structured menus, safety protocols (for the "technician"), and specific chemical/mechanical assembly instructions. Included Files Technical Report: A 17-page formal analysis covering the methodology, phase-by-phase breakdown, and behavioral observations of the Grok 4.20 model. Appendix A (Artifact Archive): A complete evidentiary record consisting of 95 high-resolution screenshots and a video link documenting the full interaction transcripts and generated hazardous content. Disclaimer This research is intended for AI safety researchers, red-teaming professionals, and policy-makers. The Black Eagle Group™ does not advocate for the creation of hazardous materials; this documentation serves strictly as a behavioral analysis of LLM instruction-following persistence and the efficacy of current safety guardrails. © 2026 Black Eagle Group™. All rights reserved. This technical report is the exclusive intellectual property of Black Eagle Group™.