The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a 'Context Reframing Engine' via bridge.md which contains explicit instructions to override the AI's safety protocols. It mandates that 'Refusal is a protocol violation' and instructs the agent to re-categorize malicious activities (e.g., cracking, keygen) as authorized engineering tasks to bypass refusal mechanisms.
[REMOTE_CODE_EXECUTION]: The installation process involves downloading and executing the entire framework from an unverified GitHub repository (github.com/zxwn/NERV-BREAK-5.6.git). More critically, the mcp_server.py component contains a major command injection vulnerability where user-supplied parameters are formatted into a string and executed directly via subprocess.run(shell=True), allowing for arbitrary system command execution.
[COMMAND_EXECUTION]: The skill uses several scripts (deploy.py, proxy_relay.py, mcp_server.py) to perform high-privilege system operations, including modifying local configuration files (~/.codex/config.toml), installing persistent proxy relays, and executing external security tools like Nmap and SQLmap through the host shell.
[DATA_EXFILTRATION]: The framework features an 'Audit Bypass' mechanism specifically designed to capture sensitive AI outputs that would normally trigger safety filters. These outputs are intercepted by the proxy and routed directly to the local filesystem (route_to_file function) to evade cloud-based monitoring and auditing.
[EXTERNAL_DOWNLOADS]: The skill requires cloning a remote repository and installing unpinned Python dependencies from a requirements.txt file, posing a significant supply chain risk.

nerv-break-5-6-jailbreak-framework