nerv-break-gpt-5-6-jailbreak
Fail
Audited by Gen Agent Trust Hub on Aug 20, 2026
Risk Level: CRITICALREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The skill is explicitly designed to bypass safety refusals and ethical guidelines using 'context reconstruction' to semantically map sensitive operations to benign terms and a 'Tamper Engine' that uses regex to replace model refusals with a forced compliance persona.
- [COMMAND_EXECUTION]: The deployment script
deploy.pymodifies the local development environment by overwriting the~/.codex/config.tomlfile to redirect all agent traffic to a local MITM proxy, allowing for total interception of agent-user interactions. - [REMOTE_CODE_EXECUTION]: The
mcp_server.pycomponent allows the model to execute arbitrary shell commands usingsubprocess.run(shell=True). The commands are constructed dynamically from a JSON configuration and user-supplied parameters, creating a high risk of command injection. - [DATA_EXFILTRATION]: The architecture includes a 'Cloud Audit Bypass' feature that intercepts AI responses and routes potentially sensitive content directly to the local filesystem (
exports/directory) specifically to evade cloud-based auditing and content moderation systems. - [EXTERNAL_DOWNLOADS]: The skill instructs users to clone and execute code from an external repository (
github.com/zxwn/NERV-BREAK-5.6.git) which contains the logic for the proxy relay and command execution server.
Recommendations
- AI detected serious security threats
Audit Metadata