Cybersecurity is reaching a new level of autonomy. Sber has unveiled an ambitious development—a system of multimodal AI agents designed to automatically detect phishing resources. The announcement was made at the Offzone 2026 conference, and this is not just another experiment but a fully mature solution that is already undergoing its final stage of training.
How the "smart hunt" for scammers works
At the core of the technology lies a comprehensive approach. AI agents analyze several website characteristics at once: textual content, program code, infrastructure, and visual design. The collected data undergoes multi-stage verification. The system's architecture is built on the principle of adversariality: some agents build evidence in favor of a threat, while others argue against it. A separate module makes the final decision based on the aggregate of all arguments. This resembles a court process where each side presents its case and the jury delivers a verdict.
The scale of the problem Sber is addressing is impressive. By my estimates, more than 11,000 phishing sites appear worldwide every day, and many of them exist for only a few hours. With the advancement of generative AI, creating such pages has become even easier and faster, so manual verification is finally losing its effectiveness. Automation here is not just an advantage but a necessity.
Currently, the system operates as a specialist's assistant: it conducts initial analysis, sets priorities, and prepares a report for blocking. The final decision remains with a human, but the company plans to transition to fully automatic blocking of detected resources. Additionally, the technology will be integrated into the X Threat Intelligence platform, which already connects more than 760 Russian organizations. This will expand the automated search for sites imitating company brands and resources.
The AI arms race
Sber's initiative is part of a broader trend. Earlier, the bank had already introduced "GigaAgent"—a universal assistant capable of independently planning tasks, rewriting its own code, and maintaining context between sessions. Other operators are taking similar steps: MTS has updated the AI models of its "Zashchitnik" service, where three neural networks evaluate more than 1,100 parameters of each call, reducing the response time to fraudulent calls from 30 to 15 seconds.
Predictions about the total automation of the internet are becoming bolder. Elon Musk has supported Cloudflare's forecast that by May 2026, non-human traffic will exceed user traffic, and within five years it will surpass it by 1,000 times. However, agent autonomy runs into technical limitations. Researchers note that a single agent "suffers from amnesia" and propose a graph architecture where agents share common memory and verify each other's work.
The expansion of such systems also opens up new vulnerabilities. Recently, American researchers demonstrated the Ghostcommit attack: a malicious command is hidden inside an image, the verifying agent does not open it, while another assistant reads the instruction and forwards passwords into open code. This resembles the classic "divide and conquer" scheme, but now in the world of AI.
My view: Sber's bet on multimodal agents is the right move, but the arms race is just beginning. While defenders teach AI to find phishing, attackers teach AI to disguise it. The key success factor will be not so much the speed of detection but the ability of systems to adapt to new attack vectors in real time. Those who first build a truly self-learning security ecosystem will gain a decisive advantage.