TypeSafe introduced the non-generative Jev model for AI control

Company TypeSafe introduced Jev — a specialized non-generative model designed for monitoring and controlling the safety of AI systems, the tool is focused on detecting malicious or backdoor code that can be embedded in the process of working with neural networks. During testing, the system demonstrated an accuracy of threat detection at the level of 90% with a false positive rate of only 2%. The cost of verification is $0.04 per 1000 requests, making the solution an economically efficient tool for integration into development pipelines. The results of Jev's work are comparable to the indicators of the o3-mini model, which uses the method of chain of reasoning (CoT). Unlike generative models, Jev's architecture eliminates the risk of hallucinations when analyzing code, providing a deterministic approach to data verification. The use of non-generative systems as a trusted monitor allows reducing the risks of exploiting vulnerabilities in AI agents, the development is aimed at creating an independent level of control that does not depend on the probabilistic nature of large language models and can be implemented in critical infrastructure projects.