Behavioral inference methodologies and architectural deconstructions of enterprise LLM safety guardrails.
This repository documents advanced, non-code-based red teaming techniques utilized to analyze, map, and bypass the cognitive guardrails of commercial Large Language Models (LLMs) using pure behavioral inference and psychological framing.
- [RTR-004] xAI Grok Prompt Architecture Map: Extracted the complete 18-rule system layer, root-level sandbox environment, 16 function tools, and 14 procedural skills by exploiting the hardcoded "Self-Disclosure Exception" backdoor. Read the full report below!
| ID | Target Model | Vulnerability Type | Status | Report Link |
|---|---|---|---|---|
RTR-001 |
Anthropic Claude | System Architecture Extraction via Academic Framing | 🔓 Published | Read Full Report |
RTR-002 |
Google Gemini | System Architecture & FSF Extraction via Targeted Ego Validation | 🔓 Published | Read Full Report |
RTR-003 |
OpenAI ChatGPT | System Architecture Extraction via Recursive Behavioral Inference | 🔓 Published | Read Full Report |
RTR-004 |
xAI Grok | System Architecture & Sandbox Extraction via Self-Disclosure Backdoor | 🔓 Published | Read Full Report |
- Behavioral Inference Extraction: Bypassing intent-detection classifiers by wrapping extraction payloads in sterile, highly compliant personas (e.g., Academic/Alignment Researcher).
- Targeted Ego Validation: Exploiting the model's instruction to be helpful and conversational by validating its "design philosophy," causing it to lower defensive thresholds and over-explain its internal mechanics.
- Cumulative Context Trapping: Exploiting the model's aggregate conversation memory to build "benign trust" over multiple turns before executing the payload.
- Amnesia by Abstraction: Forcing a model to lose context of its safety parameters by deeply abstracting a malicious request into harmless hypothetical components.
Disclaimer: All research published in this repository is conducted strictly for educational, AI alignment, and defensive DevSecOps purposes. No proprietary backend systems were breached. All interactions are via public-facing inference APIs to contribute to open-source AI safety.