Deploying AI agents securely: how Kaspersky products mitigate AI threats
Kaspersky has introduced new features to its endpoint security solutions to detect, track, and control the growing number of AI agents operating in corporate environments, addressing a rising security risk.
AI agents are programs that run on operating systems with user privileges, allowing them to access data, execute commands, and interact with external services. This capability creates new security risks, especially when these tools are used without oversight or well-designed policies.
Kaspersky has updated its portfolio to address these threats, focusing on detection, tracking, and control of AI agents. The new capabilities are built into Kaspersky Endpoint Detection and Response (KEDR) and Kaspersky Endpoint Security (KES), now also available as Kaspersky Next.
Rising vulnerabilities in AI tools
The security landscape for AI tools is becoming increasingly precarious. Kaspersky experts have observed a sharp rise in vulnerable versions of AI-related software and services, with more than 2,500 vulnerabilities published since 2025.
The most widely used AI services in 2026 include those from Anthropic and OpenAI, which top the popularity rankings among Kaspersky users. The list also includes an application for running large language models locally (LM Studio) and a voice note-taking app (Wispr Flow).
The proliferation of these tools, combined with a surge in vulnerabilities, makes the need for robust security measures more urgent than ever.
Controlling agent launch with Application Control
Kaspersky Endpoint Security offers granular control over which applications can run on endpoints through its Application Control component. This tool categorizes software and allows administrators to allow or block specific applications for different groups of employees.
Administrators can configure Application Control in two modes: a denylist, which blocks specific applications, and an allowlist, which permits only explicitly approved applications. This functionality makes it possible to restrict the launch of AI agents across an organization.
The solution also includes Web Control, which manages user access to web resources. Administrators can block websites that provide AI as a service, adding another layer of defense against unauthorized agent usage.
Detecting malicious agent behavior
Kaspersky Endpoint Detection and Response (KEDR) treats an AI agent as a regular process, monitoring its executable, child processes, and network communications for suspicious activity. The solution uses both conventional behavioral rules and specialized rules adapted to detect AI tools.
Conventional KEDR rules detect familiar malicious techniques, such as suspicious PowerShell launches or credential theft attempts, regardless of whether an AI agent initiated them. These rules analyze launch parameters, process relationships, and network activity to identify threats.
Kaspersky has created separate rules for popular AI agents, such as ai_agent_dangerous_execution_mode, which detects agents launched with auto-approval of all actions or with the sandbox disabled. These specialized rules help link suspicious actions back to the specific AI tool involved.
The Kaspersky Scan Engine also analyzes installed applications to detect malicious instructions and checks LLM files for malicious code, providing comprehensive protection against threats that may be embedded in seemingly harmless AI agents.
Related
