Generative AI has many applications in business, including generating content such as code, text, and scripts, and communicating with clients. However, the more widespread such solutions become, the more vulnerable they are to cyberattacks, with jailbreaking and prompt injections being the most common. Using these methods, criminals can hijack a model and gain access to confidential data or generate content that is typically off-limits. For instance, in 2025 hackers were able to access personal data of users by exploiting vulnerabilities in M365 Copilot and OpenAI’s Operator.
In order to protect AI against emerging vulnerabilities, the company HiveTrace (Master’s students of ITMO’s AI Talent Hub program with support from the IT company Raft) has developed an eponymous service – the first Russian AI-based protection system that deflects attacks on generative-AI web apps.
“Though there are many analogs on the international market, they aren’t adapted for the Russian market and language. Many companies in Russia are only making their first steps towards implementing AI into their business processes, but in time the demand for secure AI solutions will grow. Major tech companies do have expertise in the field of AI systems security, but their services are meant for internal use and aren’t available on the market. Meanwhile, other companies that are adopting GenAI require AI security products, too. And that’s why we’ve created HiveTrace – now any business, regardless of its size, will have access to complex protection of AI from cyberthreats,” explains Evgeny Kokuykin, CEO of HiveTrace and the head of ITMO’s AI Security Laboratory.
HiveTrace can mitigate 70% of vulnerabilities from OWASP Top 10 Risks & Mitigations, a list regularly updated by the non-profit OWASP Foundation. The service can tackle prompt injections, leaks of sensitive data or system prompts, improper input handling, excessive agency, vector- and embedding-related weaknesses, and unbounded consumption of computational resources, which causes system overload.
The new system is compatible with all popular AI models – open-source and closed ones – and can be easily made to fit a company’s security policy. Through the HiveTrace web app, security administrators can manage GenAI web apps and their security protocols. Each user query passes through the HiveTrace filter before accessing the target web app. In this process, the system analyzes the query and blocks potential threats; damaging prompt injections are detected with a 97% accuracy.
HiveTrace currently protects AI models from 7 out of 10 of the most critical exploits. Next, the company plans to increase the number of features to neutralize both existing and emerging cyberthreats. When creating new AI security solutions, the developers will rely on scientific articles and research, including those published by ITMO University researchers. The system will be of use for tech, financial, medical, and telecommunications companies. Currently, HiveTrace is prepares for pilot testing of its system.
