Claude jailbreaks written in invisible ink

What it does This repository collects jailbreak prompts and red-teaming techniques aimed at large language models, primarily Claude. The README is rendered almost entirely as Unicode variation selectors and private-use characters, making it unreadable in standard text viewers.

The interesting bit The documentation is itself a dense block of Unicode obfuscation, suggesting the project treats filter evasion not just as a topic but as a presentation style.

Key highlights

• 2,125 stars
• Explicitly focused on jailbreaking LLMs, mainly Claude
• The source text is a wall of Unicode variation selectors rather than plain Markdown

Caveats

• The README is effectively opaque in the provided source; expect to decode or reverse-engineer the content rather than read it directly
• It is unclear from the source whether the encoded payload is a prompt database, tooling, or research notes

Verdict AI safety researchers and red-teamers studying prompt-injection vectors may find this a useful curiosity. If you are looking for polished, above-board LLM tooling, keep walking.