Just AI has updated the Jay Guard protection system for Russian-speaking AI agents

Company Just AI presented updated detectors for its Jay Guard platform, designed to protect LLM applications from "jailbreak" type attacks. Developers focused on adapting the system to the specifics of the Russian language to reduce the number of false positives when filtering incoming traffic.

During the research, the team analyzed social engineering methods used to bypass AI assistant restrictions. The main problem with existing solutions was the high frequency of blocking legitimate requests in Russian. To solve this problem, engineers reworked the training samples, adding specific attack patterns characteristic of the Russian-speaking segment, which allowed increasing the accuracy of classifying malicious prompts.

The Jay Guard system functions as an intermediate security layer between the user application and the language model. It performs real-time content checking, analyzing requests for attempts at manipulation, personal data leakage, and security policy violations. The update is aimed at maintaining the relevance of protection in the face of constantly emerging new attack vectors.

Maintaining the quality of detection requires continuous retraining of models on real user traffic data. Developers note that creating effective "guardrails" is an iterative process, where the key factor is the balance between strict security and preserving the functionality of AI agents for end-users.