What companies should bear in mind when managing AI agents

AI agents are acting with increasing autonomy and can therefore – if configured incorrectly – extend their own privileges. This increases risks because traditional security controls are not designed to deal with such autonomous entities. At the same time, MCP servers allow for extensive access that is often underestimated, and the integration of Large Language Models (LLMs) increases the risk of unintended data leaks, particularly with externally hosted services.

In future, organisations will need a central register that comprehensively records all new types of identity, such as agents, bots and MCP servers. Static role models must be supplemented by dynamic, context-based permissions in order to reflect the high degree of flexibility offered by NHIs. At the same time, comprehensive documentation of all rights, actions and usage contexts will be essential. This is the only way to ensure security, transparency and governance in the long term.

Like human users, AI agents must go through a clearly defined lifecycle that governs their creation, customisation and decommissioning. Key attributes such as intended use, model type and training data form the basis for secure rights assignment. At the same time, modern identity and access management solutions enable the automated and context-sensitive management of these identities.

In future, segregation of duties (SoD) rules must also be consistently applied to autonomous agents operating in highly dynamic environments. Such agents’ access to end devices, local applications and privileged contexts must be strictly limited to prevent misuse. In addition, DLP (Data Loss Prevention) strategies and sensitivity labels ensure that sensitive data remains protected. This creates a dynamic control system that encompasses all types of identities.