AnalysisPublished 29 Oct 2026

Applying Security by Design and Cyber Immunity principles to protect AI systems | Kaspersky official blog

Source Kaspersky Security Blog

Description

This analysis discusses the challenges and security principles for protecting AI systems, particularly AI agents that autonomously perform tasks with access to critical tools. It highlights the need to treat language models as untrusted, consider all external data potentially malicious, and require human approval for irreversible actions. The approach adapts Security by Design and Cyber Immunity principles, emphasizing minimal attack surfaces, least privilege, component isolation, default deny policies, and defense in depth. The analysis also addresses risks from supply chain components and cloud model updates, as well as practical pitfalls like user fatigue and evolving vulnerabilities in AI frameworks.

Impact

The impact involves increased risk of unauthorized or unintended actions by AI agents due to their probabilistic nature and inability to reliably distinguish commands from data. Potential consequences include malicious interference, operator errors, and unstable behavior affecting production processes or data integrity. Supply chain vulnerabilities and cloud model updates add further risk. User fatigue in approving agent actions may lead to inadvertent acceptance of dangerous operations. Vulnerabilities in AI frameworks and tools can be exploited if security principles are not rigorously applied.

Defensive guidance

Security by Design principles should be integrated from the earliest design stages of AI agents. Treat language models as untrusted and external data as potentially malicious. Require mandatory human approval for irreversible actions. Minimize the trusted computing base and attack surface by limiting agent tools, permissions, and context. Enforce security policies through deterministic components such as policy monitors and sandboxes. Apply least privilege and isolate components and data processing contexts. Use default deny policies for tool calls, agent interactions, and network traffic. Implement defense in depth, including probabilistic AI guardrails as supplementary measures. Protect the supply chain by blocking or strictly limiting agent-installed components to pre-approved versions. Monitor and manage risks from cloud LLM model updates. Address user fatigue by tuning approval request frequency and clarity. Prefer compact, minimal frameworks over large, feature-rich ones to reduce vulnerabilities. Maintain allowlists carefully to avoid stale entries and loopholes.

References

kaspersky.com/blog/adopting-cyberimmunity-for-ai-agent-design-rollout/56619/ (opens in a new tab)

Community reviews

Loading reviews…

Description

AI analysisMachine-generated, updated 29 Oct 2026, 13:18 UTC

Technical details

References

Community reviews

Related threats