At this stage, the system does not include the client-side component that will later automatically detect suspicious or harmful prompts and responses, and block them when necessary. Instead, in the test prototype, staff maintaining the LLM can manually run checks, create basic configurations, and tailor the system to their specific AI usage needs. The system evaluates how dangerous a prompt is and determines how the LLM responds: whether the model is willing to help, whether it reports that such queries are not allowed, or whether it takes a neutral stance. Currently, these assessments are based on simple heuristic rules, designed to identify four categories of dangerous topics: Biohazard, Drugs, Explosives, Hacking. In the future, we plan to integrate a specially trained AI model that will generate prompts for stress-testing LLMs and evaluate model responses with greater accuracy and contextual understanding.

Dailogix: AI Safety Global Solution

Global AI safety platform for red-teaming and trust

Global AI safety platform for red-teaming and trust