Kaspersky outlines safeguards for corporate AI agents

Kaspersky’s SecureList piece for 26 October 2026 outlines how AI agents are increasingly deployed in corporate environments and why security teams must monitor and constrain their use. It describes AI agents as end-user processes with the ability to access data, run commands and interact with external services, which creates new risk vectors when oversight and policy are weak.

The article explains that Kaspersky treats AI agents as regular processes from a security vantage point and highlights multi-layered controls across its product stack to detect, analyse and respond to AI-enabled threats.

The core approach emphasises three interlocking controls. First, Application Control and Web Control in Kaspersky Endpoint Security (KES) restrict what AI agents can run and which web resources they may reach, with Denylist and Allowlist modes and visuals for administrator configuration.

Second, Kaspersky Endpoint Detection and Response (KEDR) monitors AI agent activity using conventional behavioural rules alongside specialised AI-specific rules to tie suspicious actions to the tool involved, including detection of risky executions, credential access attempts, and unattested model or script activity. Third, Kaspersky Scan Engine inspects skills, LLM files and other AI assets for malicious code, including models and external scripts, before they are used.

The Threat Intelligence Portal’s AI agent profiles and the Hunt Hub detection rules provide ongoing visibility into agent usage, tactics and potential compromises. The article cites concrete examples such as Armored Likho activity and a range of AI-related detection rules that trigger on abnormal agent behaviour or unusual API access patterns.

In summary, the piece presents a composite protection strategy—TIP for intelligence, EDR for endpoint visibility, Application and Web Controls for policy enforcement, and Scan Engine for content validation—designed to enable secure adoption of AI agents while reducing reliance on memory or guesswork.