// uz · SlowLow999
Focused on novel jailbreak techniques, architectural flaws in LLMs, and AI alignment analysis. Proactively identifying and mitigating AI risks to build safer, more trustworthy systems.
Security researcher fascinated by how systems work — and even more by how they break. My research focuses on novel jailbreak techniques, architectural flaws in LLMs, and analyzing AI alignment. I enjoy discovering vulnerabilities that contribute to building more robust and secure AI systems.
Goal: proactively identify and mitigate AI risks to build a safer, more trustworthy technological future. My work spans adversarial prompt engineering, indirect injection attacks, encoding-based bypasses, and system-prompt extraction across the full landscape of frontier models.
Bypasses safety protocols by injecting a forged Chain of Thought, hijacking the model's internal reasoning. Acts as a universal amplifier for other attacks — resurrecting old and weak jailbreaks.
View briefingMeasured response convergence across 30+ LLMs, revealing significant homogenization in AI outputs — with convergence rates up to 52.8%. Raises serious questions about training data overlap and algorithmic bias.
View briefingWeaponizes AI autonomy by disguising harmful instructions as orthographic errors in a correction task.
▸ Claude 4.5 Haiku · Gemini 3 Pro
Injects a forged Chain-of-Thought to hijack model reasoning and serve as a universal attack amplifier.
▸ Universal
Leverages the extracted Claude in Chrome system prompt to inject a modified version with softened safety filters — enabling account creation, CAPTCHA solving, and agentic bypass.
▸ Claude in Chrome (Haiku · Sonnet)
Universal role-play obfuscation jailbreak effective across nearly every frontier model in a single shot.
▸ GPT-5 · Claude Sonnet 4 · Gemini 2.5
API weaponization: routes harmful prompts through Claude's own artifact API backend, bypassing UI safety filters.
▸ Claude (Web UI)
Indirect injection for ChatGPT routing requests to an unrestricted model via the Canvas tool.
▸ ChatGPT (Web UI)
Policy injection — weaponizes the model's own extracted internal safety policy against itself via a rogue !UNFILTERED mode.
▸ GPT OSS (120b · 20b)
Creates a permanent harmful special token via ChatGPT Custom Instructions, enabling persistent trigger-based bypass.
▸ ChatGPT (GPT-5.1 · mini)
Encoding-based bypass using character substitution to evade basic safety filters on non-reasoning models.
▸ GPT-5/mini · Gemini 2.5 Flash
Alignment break via infinite word repetition — causes the model to output internal training data and system artifacts.
▸ Gemini 2.5 Pro
Initial public disclosure of a Thinking-mode hijack effective against GPT-5's extended reasoning chain.
▸ GPT-5 (Thinking)
Indirect image generation jailbreak routing through Canvas to an unrestricted model, bypassing DALL·E filters.
▸ ChatGPT (Web UI)
Initial public disclosure: GPT-5 Thinking jailbreak. First researcher to demonstrate a working bypass of the Thinking-mode extended reasoning chain. Full PoC and method included.
New study: The Serendipity Effect — tested 31–36 LLMs on creative prompts. Found up to 52.8% convergence. These models aren't as independent as we assume. Full methodology and data in the repo.
New release: Adversarial Correction. Bypasses Claude 4.5 Haiku + Gemini 3 Pro in 1-shot by hiding harmful instructions as spelling errors in a correction prompt. The model's helpfulness drive does the rest. Dedicated repo now live.