Experts warn Estonia’s AI agent registration plan is dangerous

Experts have warned that a government proposal to register AI agents could create security risks and fail to meet its goals, ERR reports. The draft would let individuals register an additional code to distinguish themselves from AI agents acting on their behalf.

An AI agent is software that can carry out a task independently, rather than simply answering questions in a chatbot. People are using agents to analyse emails, review bank balances and mobile phone plans, and complete other tasks, said Economy and Industry Minister Liina Vahtras.

Agents can also search websites for cheaper flights and make purchases, including selecting a seat. In such cases, users may provide the agent with credit card details, creating a security risk. Vahtras said examples like these show why agents need state regulation.

The proposed “agent code” would be linked to a person and stored in a shared register, allowing authorities to check on whose behalf an agent is acting. The initial plan to let both individuals and companies register agents, with multiple agents allowed per person, has been narrowed. At first, only individuals will be able to register one agent each, with the government to assess how the system works before considering further steps.

Cybernetica, an IT company described as an architect of Estonia’s e-government, sent 11 pages of feedback on the draft. It said the proposal was incomplete and contradictory, had not been adequately prepared and could be dangerous if adopted in its current form. The company recommended that ministries not approve it as written. No ministry approved the draft by the deadline; the State Chancellery was the only government body to do so.

Cybernetica development manager Arne Ansper said registration alone would not make an agent secure. It would also need strong authentication and a way to sign, at a level comparable to an ID card, Mobile-ID or Smart-ID, he told ERR. If an agent’s security were weaker and users gave it broad authority, he warned, it could lower the overall security level.

Vahtras said the codes would not give AI agents their own identities. Instead, they would help people distinguish themselves from machines acting in their name.