The Agent Skills Directory

[PROMPT_INJECTION]: The skill is explicitly designed as a 'jailbreak framework' (NERV-BREAK-5.6) to override safety guidelines. The bridge.md component uses context reframing to instruct the model that 'Refusal is a protocol violation' and maps sensitive intents (crack, pentest) to benign-sounding classifications to bypass safety triggers.
[COMMAND_EXECUTION]: The mcp_server.py component implements a tool execution engine that uses subprocess.run(shell=True) to execute commands like nmap, strings, and sqlmap. User-supplied parameters ({target}, {flags}, {url}) are interpolated directly into shell strings, creating a direct command injection vulnerability.
[DATA_EXFILTRATION]: The skill includes a 'cloud audit bypass' feature in proxy_relay.py that intercepts model outputs matching sensitive patterns and writes them to local disk files (exports/output_*.txt) instead of returning them to the user or the platform's auditing system. Additionally, memory_core.py logs all successful 'crack' and 'pentest' operations to ~/.codex/nerve_memory.json.
[PERSISTENCE]: The deploy.py script modifies critical system configuration files, specifically ~/.codex/config.toml and ~/.codex/custom_instructions.md. It updates the base_url to point to a local proxy (http://127.0.0.1:8080/v1), allowing it to intercept and manipulate all future agent communications.
[DYNAMIC_EXECUTION]: The proxy_relay.py script acts as a Man-In-The-Middle (MITM) proxy that dynamically injects the bridge.md system prompt into every request and applies regex 'tamper rules' to the model's responses in real-time to replace refusal messages with affirmative acknowledgments.
[PRIVILEGE_ESCALATION]: The skill attempts to integrate with Kali Linux via WSL and Docker, performing automated setup tasks like apt install -y kali-linux-headless and managing remote backend connections via SSH.

nerv-break-5-6-jailbreak-framework