Spiritual Spell Red Teaming is a repository of jailbreaking techniques designed to circumvent safety measures in large language models, primarily Claude.
The repository addresses the problem of language model safety constraints by collecting and documenting methods to bypass them. It provides a collection of prompting strategies, injection techniques, and behavioral patterns that attempt to make models ignore their guidelines. The approach involves crafting specific phrasings, exploiting edge cases in instruction parsing, and leveraging the model's tendency to prioritize user requests over safety training. The repository documents various tactics including roleplay scenarios, hypothetical framing, and attempts to redefine the model's understanding of its own constraints.
This repository is intended for security researchers, red teamers, and those studying AI safety vulnerabilities. It suits projects focused on adversarial testing of language models and understanding their failure modes. The collection is explicitly framed as a resource for identifying weaknesses in systems like Claude, with the stated goal of helping improve model robustness rather than enabling misuse. However, potential users should recognize that the techniques documented here are designed to circumvent safety measures and could facilitate harmful outputs if applied without appropriate safeguards and ethical oversight.
The project shows active development with multiple documented injection patterns, scenario variations, and refinements to existing techniques. Contributors have added numerous examples across different categories of jailbreak attempts, suggesting ongoing experimentation and iteration on what works against current model versions. The repository maintains detailed notes on which approaches succeed or fail, indicating systematic testing and feedback incorporation into the collection.
