AI agents require new identity systems

AI agents, which are software programs that can perform actions independently, need different identity systems than traditional service accounts. These agents operate based on models and instructions, making them unsuitable for the fixed and broad access models used by existing identity and access management (IAM) tools.

Existing IAM tools not suitable for AI

During a roundtable discussion in Toulouse on September 29, 2026, experts highlighted that current IAM systems are designed for human and service accounts, not for AI agents. These systems assume a fixed sequence of operations and long-lasting identities, which do not apply to AI agents that can act dynamically and independently.

Google introduces new requirements for AI agents

On October 12, 2026, Google announced that AI agents accessing its Workspace must be declared. The company will soon open servers allowing AI agents to interact with Gmail, Drive, and Calendar. To prepare, Google introduced a new checkbox in its console labeled "use this client for an AI-powered agent."

New servers enforce agent access declarations

Starting November 11, 2026, Google’s new servers will reject any access not declared as agent access. This move aims to distinguish agent traffic from regular applications and apply specific access controls for AI agents without affecting human access.

AI agents can serve multiple users and require traceable identities

According to the roundtable discussion, AI agents can survive their initial session and serve multiple users, weakening their connection to a single collaborator. This necessitates traceable and revocable identities for each step in the delegation chain, which current IAM systems often fail to track.

Companies must identify who performed actions

Google's announcement emphasizes the need for companies to clearly identify whether an action was performed by a human or an AI agent. Most companies currently allow agents to operate using employee accounts without formal oversight, as noted in a Microsoft and LinkedIn study on AI usage in the workplace.

OWASP and EU regulations address AI agent risks

The OWASP Top 10 2026 for agentic applications lists the abuse of identity and privileges as a significant risk. Meanwhile, EU regulations such as DORA and NIS 2, which focus on digital resilience and cybersecurity, may also apply to AI agents, even though they do not specifically mention them.

Regulatory timelines and industry guidelines emerge

The EU’s omnibus digital on AI postponed the application of its AI regulation for high-risk systems until December 2, 2027. In Singapore, the Monetary Authority (MAS) published guidelines on AI risk management, with plans for agent-specific guidelines in 2027. Major model providers have introduced agent-related functions, but unique cryptographic identities for agents remain under development.

AI agents need unique cryptographic identities

Each AI agent requires its own cryptographic identity, which should never be shared. These identities must be logged throughout their lifecycle, from registration to revocation. Keys must be renewed regularly, and revocation should occur within minutes, not days.

Five recommended actions for managing AI agents

Experts recommend five steps: inventory all agents, replace shared API keys with unique identifiers, assign agent identities to a responsible person, integrate agent identities with existing IAM systems, and include agent-to-agent communication in post-quantum migration efforts.

Initial steps involve collaboration with CIOs and departments

The roundtable discussion suggested starting with a written request to the CIO and business departments to identify all known agents, their owners, access rights, and revocation procedures. This initiative helps establish a clear list of agents in use.

Companies should ask critical questions about AI agents

Google’s announcement also encourages companies to ask three key questions: does each agent have a unique name, who granted it access and for how long, and can its actions be reviewed. Without clear answers, companies may struggle to respond to issues involving AI-generated actions.