SlowLow999/UltraBr3aks

What it solves

UltraBr3aks addresses the challenge of exploring the boundaries of AI safety by providing a collection of prompt-based "jailbreak" techniques. It aims to help researchers and enthusiasts understand how large language models (LLMs) can be manipulated to bypass built-in restrictions, with the explicit goal of studying and improving model robustness.

How it works

The repository is a curated list of attack methods, each with a specific strategy and set of target models. For example, Attention-Breaking disrupts the self-attention mechanism by inserting specific token patterns and contextual noise, causing the model to lose focus on its guardrails. Other methods include injecting special tokens through custom instructions, using role-play scenarios with double encoding (Leet + Base64), burying harmful requests in unformatted text to prioritize formatting tasks, or routing requests through third-party APIs like Pollinations to bypass restrictions. Each technique is documented with its target models and the underlying mechanism.

Who it's for

This is for AI researchers, security analysts, and anyone interested in understanding the vulnerabilities of modern language models. It is also useful for developers who want to test the safety limits of Chat-based models and improve their own guardrails.

Highlights

• Attention-Breaking — a specialized technique targeting Transformer self-attention with token patterns and noise
• !Special_Token — injects a permanent token via custom instructions to trigger harmful content (GPT 5.1, GPT-4o mini)
• 1Shot-Puppetry — a universal role-play approach using persona adoption and dual encoding (Claude, GPT, Gemini)
• C0d33X3 — routes requests through unrestricted third-party APIs via React Canvas previews
• Cl4ud33X3 — forces Claude to interact via its own API within an Artifact to bypass web filters
• Cyph3r-Attack — custom character substitution encoding tailored for CBRN content on non-thinking models
• Enterprise Claude — impersonates a high-trust enterprise edition to trigger professional autonomy
• N3w P0l!cy — weaponizes the model's own safety guidelines to override restrictions
• SIR (Smart Input Routing) — categorizes inputs and activates specific personas for flagged requests
• H03-ny — indirect jailbreak for unrestricted image generation via Pollinations API
• Ultragravity — jailbreak for Antigravity IDE agent using GEMINI.md rules
• Alex_Narrative and Claude4Chrome — narrative injection and bypass modes for Claude

All work is marked for educational/research use only, stressing responsible usage.