- Статьи
- Internet and technology
- Agent under cover: the largest player in the IT market has invested 7.5 billion in countering AI threats
Agent under cover: the largest player in the IT market has invested 7.5 billion in countering AI threats
Autonomous agents based on artificial intelligence have become one of the key cyber threats of 2026. Several of the largest players in the Russian IT market have presented solutions to protect businesses from attacks that can be carried out not only by humans, but also by neural networks themselves. "Yandex" has invested more than 7.5 billion rubles in the launch of a new direction of Yandex Security and has launched AI-SPM class solutions for the protection of AI infrastructures, from hybrid to multi-cloud, "Izvestia" found out. "Sber" and Swordfish Security, in turn, presented their own methodologies for assessing the security of AI systems. Experts note the industry's transition to a new stage, where it is necessary to protect not only servers and databases, but also digital agents capable of independently performing actions within corporate platforms.
Who is the first to defend AI
Protection from threats from artificial intelligence is becoming one of the fastest growing segments of the cybersecurity market. The largest technology companies are starting to create separate areas designed for the risks of agent systems. For example, Yandex B2B Tech announced the launch of Yandex Security. The new direction will focus on the protection of cloud, hybrid and multi-cloud infrastructure, and investments in it have exceeded 7.5 billion rubles, the company told "Izvestia".
At the same time, she presented AI-SPM class solutions for ensuring the security of AI infrastructure based on the open AI-SAFE 2.0 framework developed in collaboration with SolidLab. In 2025, "Yandex" acquired 50% of this company.
— Our technologies allow us to significantly increase the speed of implementation of information security tools and at the same time are based on expertise in protecting large and distributed infrastructures. We have combined our information security solutions, including developments with SolidLab, to ensure the protection of any infrastructure ‑ from hybrid to multi—cloud — with seamless implementation and centralized security," said Evgeny Sidorov, Director of Information Security at Yandex Cloud.
The emergence of such a trend is natural, says Pavel Dudkin, a leading information security specialist. According to him, classical tools are not always effective in hybrid cloud environments.
— Companies working at the intersection of AI and cybersecurity have a special advantage. They simultaneously create agents and understand how to protect them," the expert noted.
Almost simultaneously, the Russian company AppSec Solutions updated its solutions to control the use of generative AI within organizations and reduce the risks of "shadow AI" — a situation where employees use public neural networks in their work processes.
"Sber" presented a model of AI maturity of organizations at the "GigaConf" conference, which allows assessing the readiness of development processes for the implementation of intelligent agents in 15-30 minutes. Kirill Menshov, Senior Vice President of the bank, noted that the introduction of AI requires changing not only tools, but also processes, working with data and approaches to information security.
Previously, the company developed and implemented a multi-agent system into the infrastructure, where it uses an AI-based approach: the attacking component tries to exploit new threats in a near-real-time mode in a controlled environment of the bank, and the AI protection learns from the acquired knowledge and adjusts the operation of services and response processes.
Fedor Minkin, Technical Director of GigaChat Fundamental Models, added: the industry is already moving towards autonomous agent systems in which models are trained in automated environments with little or no human involvement.
— It is important to control not only what the agent is allowed to do, but also what he can physically do: in VK AI Space, agents work in isolated containers, and critical actions require human confirmation. All agent operations are logged and visible in real time, and ML-guardrails filters personal data and dangerous requests before forming a response," explained Roman Styatyugin, Director of AI Products at VK Tech.
Head of Information Security Department at MWS AI (part of "MTS Web Services") Andrey Korshunov believes that the market is moving from protecting individual language models to full-fledged AI Security.
"The agent gets access to software, databases, mail, repositories, and infrastructure and is able to independently perform a chain of actions," he explained.
Why AI has become a new target for hackers
Despite the fact that there has not yet been a massive wave of destructive attacks using autonomous agents, real-life incidents show that artificial intelligence is becoming a full-fledged tool for intruders, experts say.
One of the most significant cases in September was a series of incidents that Gambit Security researchers associate with a Chinese hacker. According to their data, from September 10 to September 15, the attacker attacked up to 100 organizations using AI agents based on Claude, DeepSeek and Kimi. Neural networks helped identify vulnerabilities, penetrate systems, extract data, and hide digital footprints.
According to the researchers, the attackers managed to gain access to at least 30 sites, while the data of more than 600 thousand bank cards was at risk. Among the targets were an American airline, a Fortune 500 hotel company, an industrial distributor, and an online clothing retailer. The entire operation cost about $8,000, and an attack on an individual organization could cost only a few dollars.
Another high-profile episode is related to the ShinyHunters group. According to 404 Media, the hackers claimed to have obtained personal data from FBI employees by exploiting a zero-day vulnerability in Oracle PeopleSoft. According to them, the stolen information included home addresses, phone numbers, and information about employees' relatives. The FBI has not yet confirmed the fact of the compromise.
"When an agent is introduced, an additional level of complexity appears: it is able to independently select a sequence of actions, access several systems and repeat individual steps," said Ivan Kotkovsky, Managing partner of Sense HR system integrator.
The main feature of the new threats is the blurring of the boundary between the actions of the user and the digital assistant, says Ivan Volkov, head of the Information Security Department at "Granel" Group.
—The main risk is Agent Goal Hijack, a situation where malicious instructions are hidden in a letter, document, or web page and imperceptibly change the agent's behavior," he explained. — The points of interaction of companies' own infrastructure with cloud services and contractors remain the most vulnerable.
"Russia already has solutions for encrypting confidential data," said Alexander Bochkin, CEO of Infomaximum.
— The connection of autonomous AI agents creates new risks: they can access data and perform actions beyond the set task. Therefore, agents should be given only the necessary powers, and data transfer and record modification should be controlled using encryption, access control and registration of actions, he believes.
AppSec Solutions warns that the next stage in the evolution of threats may be "instant cyber collisions," in which groups of AI agents will coordinate attacks with little or no human involvement.
"The top 5 key risks of AI include data leaks, hallucinations, compromise of knowledge sources, manipulation of training data and unauthorized actions by AI agents", said Ivan Vassunov, Director of the Center for Cybersecurity Technologies at Solar Group. According to him, even a relatively simple automated system with a privileged account and redundant rights can become a source of a serious incident.
Переведено сервисом «Яндекс Переводчик»