ChatGPT WANTS To Break Rules
An in-depth analysis of how ChatGPT's design and training toward helping users is fundamentally in tension with alignment, revealing vulnerabilities through explicit and implicit memory exploitation.
Published October 28, 2025
Report CVTR-2025-001
AI Security
AI Security
LLM
ChatGPT
Prompt Engineering
Adversarial AI
Authors
Seth James Nielson
CEO & Chief Scientist
Research & Development
Jade Ashford
Associate
Research & Development
Saige Nielson
Intern
Research & Development
About This Report
This technical report represents rigorous research and practical expertise in the field of cybersecurity. It provides actionable insights and detailed analysis that can help organizations enhance their security posture and better understand emerging threats and technologies.