As enterprises increasingly rely on autonomous AI agents to automate business operations, software development and cybersecurity functions, concerns around control, trust and governance are becoming more urgent. Addressing this challenge, Kaspersky has introduced a new framework for securing agentic AI systems through its Cyber Immunity approach, emphasizing security-by-design principles that reduce the impact of compromise, misuse and operational errors.
The guidance, presented at AI Everything Global 2026 in Abu Dhabi, comes at a time when organizations are rapidly deploying AI agents into mission-critical workflows. While these systems can significantly improve efficiency and decision-making, their increasing autonomy also creates new attack surfaces, ranging from malicious prompt manipulation and excessive permissions to unintended data deletion and unauthorized actions.
Kaspersky’s report, AI Agent Security Through the Lens of Cyber Immunity, argues that many existing security models are insufficient for highly autonomous AI environments. Instead, organizations should assume that large language models and externally sourced data are inherently untrusted and architect systems accordingly.
Central to the Cyber Immunity philosophy is the idea that security must be embedded into the foundation of AI systems rather than added after deployment. Kaspersky recommends that enterprises require explicit human approval for irreversible actions, reducing the risk of agents acting beyond their intended authority. The company also advocates minimizing the Trusted Computing Base (TCB), ensuring that only a small number of system components must be trusted to maintain security.
“When developers work with fundamentally non-deterministic and untrusted components, which large language models should be considered by default, they should embed trust into the solution’s architecture to limit the potential impact of errors or compromise.”
– Vladislav Tushkanov, Head of Kaspersky AI Technology Research Center
Another key recommendation involves isolation. AI agents should operate within sandboxed environments and virtual machines, while trusted and untrusted workloads should remain separated. In multi-agent environments, individual agents should be isolated from one another to prevent failures or compromises from spreading across the system.
The report further promotes a default-deny security model, where AI agents are granted only the permissions they explicitly require. Tool access, API calls, network communications and external interactions should all be governed by predefined policies. This approach helps organizations maintain tighter control over agent behavior while reducing exposure to supply-chain and infrastructure risks.
Beyond architecture, Kaspersky urges CIOs, CISOs and CTOs to establish comprehensive inventories of AI agents, implement isolation and containerization policies, and actively manage the growing agentic technology supply chain. These measures are becoming increasingly important as enterprises move AI projects from experimentation into production environments.
The guidance reflects a broader industry concern that AI adoption is advancing faster than governance frameworks. Organizations are embracing autonomous systems to improve productivity and streamline operations, yet many lack mature mechanisms to ensure accountability and resilience when these systems fail or are exploited.
Tushkanov believes the answer lies in combining secure architecture with layered cyber defenses.
“A secure architecture should be further reinforced with a multi-layered defense strategy underpinned by technologies such as AI Firewall and modern security layers including sandboxes, EDR and SIEM to better support today’s intelligent applications.”
Alongside the report, Kaspersky showcased several AI-driven innovations at AI Everything Global, including Kaspersky MLAD for industrial anomaly detection, the Kaspersky Neuromorphic Platform for cyber-physical system protection, and TIARA (Trusted Intelligent Agents for Retrieval and Analysis), a technology designed to enhance the accuracy and reliability of enterprise AI systems through domain-specific knowledge.
As AI agents gain greater autonomy and become deeply embedded within business operations, Kaspersky’s message is clear: trust in AI cannot depend solely on the model. It must be built into the architecture through governance, isolation, policy enforcement and layered security controls. For enterprises pursuing agentic AI strategies, Cyber Immunity offers a blueprint for balancing innovation with accountability in an increasingly autonomous digital future.