Kaspersky publishes AI-agent security guidance built around untrusted-by-default models, human confirmation and default-deny tool controls

What changed

Kaspersky published 'AI agent security through the lens of Cyber Immunity' on October 7, applying secure-by-design Cyber Immunity principles to autonomous agents. The guidance recommends treating LLMs and externally supplied data as untrusted by default, requiring explicit human confirmation for irreversible actions, minimizing the Trusted Computing Base, isolating agents and contexts with sandboxes or virtual machines, and applying default-deny policies to tool calls, arguments, network access and the agentic supply chain. It also advises organizations to inventory agents and formalize isolation and supply-chain controls. This is security guidance and architecture advice, not independent proof that any one approach eliminates agent risk.