Verify local AI agent defenses in a controlled environment
You can directly test how well a DeepSeek Harness installed on your computer blocks malicious instructions. This tool separates the model's responses from the system's blocking mechanisms, allowing you to accurately identify actual vulnerabilities. It is a dedicated plugin required for performing security assessments safely in a local environment.
Discontinued maintenance and license change
Due to personal circumstances, the author will stop updating the GitHub repository after October 26, 2026. Subsequent versions will be distributed only via a Telegram channel, and the existing MIT license will be changed to a proprietary license that restricts commercial use. Use after this date is subject to the new license terms.
Separating model and system responses
DeepSeek Harness does not simply ask the model questions; it has multiple layers of defense mechanisms, such as modifying prompts during execution and going through approval procedures. This plugin distinguishes between these system-level interventions and the model's own refusals, clearly showing at which stage a request was blocked. This provides a basis for determining whether the default security policy is excessively strict or loose.
Scope of application in local environments
This tool only affects the official DeepSeek Harness package (@deepseek-ai) and configuration files installed on the user's local computer. It does not include scanning or attack capabilities for external network servers or other people's systems, and all patching and injection processes occur only within the local file system. The web version (dsh web) and the desktop version (EXE) must be installed and applied independently.
Essential checks before application
Users must run this tool only in local environments they manage or on offline test targets. Access to servers or websites in mainland China, Hong Kong, or Macau is strictly prohibited without prior registration. Additionally, if you intend to resell commercially or sell educational courses, you must purchase a separate commercial license, while simple personal use is permitted. A Node.js 18 or higher environment is required, and Node.js 22 or higher is recommended to resolve window flickering issues on Windows.