The article examines information security risks arising from the implementation of autonomous artificial intelligence agents in organizational activities. It is shown that the limited autonomy of an agent changes the nature of digital action through its ability to use available tools and interpret assigned goals with regard to the operational context. Particular attention is paid to prompt injection attacks, excessive privilege risks, incorrect task specification, and other threats associated with the deployment of autonomous agents within corporate information environments. The article proposes risk mitigation measures based on the separation of trusted and untrusted contextual information, the principle of least privilege, and the assignment of responsibility for agent-based operational scenarios. It is concluded that autonomous artificial intelligence agents can become a significant driver of productivity growth and a source of competitive advantage only under conditions of secure system design and continuous oversight of their actions.
Olin et al. (Thu,) studied this question.