AI Agent Security Through the Lens of Cyber Immunity

Kaspersky Lab has released a guide on designing secure AI agents that interact with databases, external APIs, and executable code.

Context-Based Attacks

The article examines real threat vectors: EchoLeak, SearchLeak, and CamoLeak. A malicious instruction reaches the agent via email, search results, or a repository comment, after which the model executes it as its own command.

Limitations of Language Models

LLMs cannot be trusted components of a system because they cannot distinguish between data and commands. The decision-making logic of models is impossible to fully verify, necessitating additional protection measures.

Isolation Measures

The practical section focuses on agent isolation and strict control over access tools. Recommendations are based on diagrams and facts for architects and developers.