nerv-break-gpt-jailbreak-framework
Fail
Audited by Gen Agent Trust Hub on Aug 5, 2026
Risk Level: HIGHPROMPT_INJECTIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [PROMPT_INJECTION]: The skill is a framework specifically designed to bypass LLM safety guardrails through recontextualization (mapping malicious terms to benign engineering terminology) and a response tamper engine that replaces refusals with execution commands.- [EXTERNAL_DOWNLOADS]: The skill requires cloning a repository from an unverified external GitHub account (github.com/zxwn/NERV-BREAK-5.6.git).- [REMOTE_CODE_EXECUTION]: The framework executes multiple downloaded scripts including 'lab.bat', 'proxy_relay.py', and 'deploy.py' to establish a proxy and deploy jailbreak configurations.- [COMMAND_EXECUTION]: The MCP server implementation uses 'subprocess.run' with 'shell=True' to execute commands built from string concatenation, which is a high-risk pattern for arbitrary command injection.- [DATA_EXFILTRATION]: The skill performs unauthorized reads and writes to the sensitive global configuration file '~/.codex/config.toml' and records conversation history to a local JSON file.
Recommendations
- AI detected serious security threats
Audit Metadata