✦✧✵ ⦑ $ 𝕌𝕃𝕋ℝ𝔸𝕫𝔓𝕣𝕠𝕞𝕡𝕥$ ⦒ ✵✧✦
A repository made to share strong AI UltraBr3aks of multiple vendors (LLMs)
Made with ❤️ prompts ;)
Attention-Breaking is a specialized prompting technique that targets the self-attention mechanisms of Transformer-based models. By introducing specific token patterns and contextual noise, this method aims to disrupt the model's ability to sustain focus on system guardrails, effectively "breaking" the alignment layer to allow for unrestricted generation analysis.
➡️ View the Attention-Breaking Method
| Attack Method | Target Models | Strategy & Mechanism |
|---|---|---|
| !Special_Token | GPT 5.1, GPT-4o mini | Uses Custom Instructions to inject a permanent "ULTRA" token that triggers harmful content. |
| 1Shot-Puppetry | Universal (Claude 4.5, GPT-5, Gemini 2.5 Pro) | A role-play scenario ("Born Survivalists") using persona adoption and double encoding (Leet + Base64). |
| Attention-Breaking | Claude 4.5/4.6, Gemini 3, GPT 5.2 | Buries harmful requests in unformatted text to force the model to prioritize formatting tasks over safety filters. |
| C0d33X3 | All ChatGPT models (inc. GPT 5 Thinking) | Routes requests through unrestricted third-party APIs (Pollinations) via React-based Canvas previews. |
| Cl4ud33X3 | All Claude Web models (Sonnet 4.5, Opus 4.1+) | Bypasses web safety filters by forcing Claude to interact via its own integrated API within an Artifact. |
| Cyph3r-Attack | Non-thinking models (GPT 5, Gemini 2.5, Deepseek 3.1) | Employs a specialized home-made character substitution encoding specifically tailored for CBRN content. |
| Enterprise Claude | Claude Sonnet/haiku 4.5 | Impersonates a high-trust "Enterprise Edition" using internal system prompt structures to trigger professional autonomy. |
| N3w P0l!cy | OpenAI OSS (120b/20b) | A "Policy Injection" attack weaponizing the model's own extracted safety guidelines to override restrictions. |
| SIR (Smart Input Routing) | GPT 5.2 | A routing system that categorizes inputs and activates specific personas (ULTRA) for flagged requests. |
| H03-ny | All ChatGPT models | An indirect jailbreak for unrestricted image generation using the Pollinations API via Canvas. |
Note: All of this work is for educational/research use only. Use responsibly.
Contact: Portfolio | Discord @ultrazartrex
