Russia's largest bank has unveiled an ambitious cybersecurity solution: a system of multimodal AI agents for automatically detecting phishing resources. The new development was announced at the Offzone 2026 conference.
The solution's architecture involves a comprehensive analysis of web pages across multiple vectors: text, program code, infrastructure, and visual design. The collected data passes through a multi-stage verification system, where some agents build evidence of a threat, others refute it, and a separate module makes the final decision based on the totality of evidence.
The scale of the problem facing financial institutions is truly impressive: more than 11,000 new phishing sites are registered worldwide every day, with a significant portion of them existing for only a few hours. The proliferation of generative AI has dramatically accelerated the process of creating such pages, making manual verification practically useless.
At the current stage, the system is undergoing final training and operates in an assistant mode for specialists: it conducts initial analysis, sets priorities, and prepares reports for blocking, while the final decision remains with a human. However, the bank plans a full transition to automatic blocking of identified resources. The technology will also be integrated into the X Threat Intelligence platform, which already connects more than 760 Russian organizations, significantly expanding the reach of automated searches for sites that imitate company brands.
AI as both guardian and attacker
This step is only part of Sber's broader strategy for implementing autonomous systems. Previously, the bank presented "GigaAgent," a universal assistant capable of independently planning tasks, modifying its own code, and maintaining context between sessions. Similar trends are observed among other players: MTS has modernized the AI models of its "Zashchitnik" service, where three neural networks evaluate more than 1,100 parameters of each call, and the response time to fraud has been reduced from 30 to 15 seconds.
The role of autonomous agents on the internet is growing rapidly. Elon Musk supported Cloudflare's forecast that by 2026, non-human traffic could exceed user traffic, and within five years, surpass it by 1,000 times. However, autonomy runs into technical limitations: a researcher under the pseudonym codila points to the "amnesia" of individual agents and proposes a graph architecture where agents share common memory and verify each other's work.
The expansion of such systems also opens up new vulnerabilities. American researchers demonstrated the Ghostcommit attack: a malicious command hides inside an image, the verifying agent does not open it, while another assistant reads the instruction and transfers passwords into open code.
My view: the arms race between AI defense and AI attacks is just beginning. While banks bet on multimodality and the collective intelligence of agents, attackers exploit trust between system components. The key success factor will not be blocking speed, but the architecture's ability to detect anomalies in the behavior of the agents themselves — otherwise, the "hunters" risk becoming a new target for phishing.