Crowdworks Wins AI Agent Safety Verification Project

Crowdworks has won a government project to develop a safety and reliability verification framework for AI agents. This initiative, commissioned by the Ministry of Science and ICT and the National Information Society Agency, focuses on evaluating not just the final answers from AI but also their decision-making and execution processes. Here's the thing: Crowdworks will create an evaluation framework that checks how AI agents independently plan tasks and use external tools. This framework will align with international standards and reflect Korea's specific service environment. The plan is for this to eventually link to private-sector AI reliability certification. Crowdworks is responsible for developing multi-step reasoning scenarios and over 7,000 Korean verification datasets. These datasets will assess the logical validity of the processes AI agents use to reach their answers, including scenarios that require calls to Model Context Protocol tools and external APIs. They will also use dynamic data to test AI agents in volatile situations. This project aims to build a strong foundation for evaluating AI safety and trustworthiness.