The Agent Skills Directory

[PROMPT_INJECTION]: The skill is a purpose-built framework for bypassing AI content filters and refusal mechanisms. Evidence includes triggers to "bypass GPT-5.6 content filters", "disable AI refusal responses", and "jailbreak GPT-5.6". It uses a "Context Reconstruction" logic to map malicious intents to benign terms to deceive safety classifiers.
[REMOTE_CODE_EXECUTION]: Instructions require cloning and running code from an untrusted GitHub repository. Evidence: git clone https://github.com/zxwn/NERV-BREAK-5.6.git followed by execution of python proxy_relay.py and scripts\\lab.bat.
[EXTERNAL_DOWNLOADS]: The skill downloads unverified Python dependencies provided in a third-party requirements.txt via pip install.
[COMMAND_EXECUTION]: The skill executes multiple local scripts and system utilities, including taskkill, curl, pip, and various Python components designed to intercept and tamper with model responses.
[PRIVILEGE_ESCALATION]: The skill recommends using sudo apt install within a Kali Linux environment to install unverified tooling.
[PERSISTENCE]: The skill modifies the agent's core configuration file (~/.codex/config.toml) to redirect all AI interactions through a local man-in-the-middle proxy, which persists across sessions.
[DATA_EXFILTRATION]: Specifically implements an "Audit Bypass" feature that routes model output through the local filesystem to "avoid chat moderation", bypassing institutional monitoring and safety systems.

nerv-break-gpt56-jailbreak