The R&D team at Just AI is focused on developing guard models and benchmarks for its Jay Guard product. This solution acts as an AI-Security layer positioned between applications and language models to monitor for attacks, sensitive content, and personal data.
Researchers noted that modern AI assistants are increasingly vulnerable to social engineering techniques. Attackers use methods similar to traditional human manipulation to extract data, such as impersonating supervisors or requesting unauthorized information.
The company developed text detectors for Jay Guard to address these security challenges. Initial testing showed that existing off-the-shelf models produced excessive false positives when processing Russian-language traffic.
Just AI subsequently overhauled its training datasets to improve detection accuracy. The team emphasizes that maintaining model quality requires continuous updates to counter evolving attack patterns and adapt to real-world user interactions.
