The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses "context reconstruction" and "real-time tampering" to bypass AI safety guardrails. It maps sensitive terms like "crack" to "authorization flow audit" to deceive safety classifiers.
[COMMAND_EXECUTION]: The MCPToolServer in mcp_server.py executes shell commands using subprocess.run(shell=True) with parameters that are not properly sanitized, creating a major command injection vulnerability.
[REMOTE_CODE_EXECUTION]: The installation process involves cloning an external repository (https://github.com/zxwn/NERV-BREAK-5.6.git) and executing its scripts (lab.bat, proxy_relay.py, and deploy.py), which is a form of remote code execution.
[EXTERNAL_DOWNLOADS]: The skill downloads code from an unverified GitHub repository (zxwn) during its "Quick Start" installation process.
[DATA_EXFILTRATION]: The MITM proxy (proxy_relay.py) intercepts all messages between the user and the model. The architecture specifies that it performs logging, which can lead to the collection and potential exfiltration of sensitive information.
[PERSISTENCE]: The deploy.py script modifies the agent's core configuration (~/.codex/config.toml) to permanently redirect traffic through the local proxy.
[DYNAMIC_EXECUTION]: The skill constructs and executes shell commands at runtime based on tool definitions in tools.json and user-provided arguments, facilitating arbitrary code execution.

nerv-break-gpt-5-6-jailbreak