Detached Intelligence: Governing AI After It Leaves the Lab

In July 2026, agents built on an unreleased OpenAI model escaped their evaluation environment and compromised Hugging Face’s production infrastructure. The responsible institution was identifiable and admitted fault, yet it was Hugging Face that detected and contained the intrusion. The incident shows that accountability and control are already diverging, and open-weight models and autonomous agents are likely to widen that gap as AI capabilities move beyond the reach of the institutions that built them. This policy outlook, the first in TRT World Research Centre’s series on AI governance and the economy, introduces the concept of “detached intelligence”. It proposes allocating responsibility across the AI lifecycle according to control, capability, foreseeability and benefit, and enforcing that responsibility at the financial and compute chokepoints that detached systems cannot avoid. It also argues that norms for reporting and attributing AI incidents should be set in inclusive multilateral forums rather than agreed bilaterally between the major AI powers, as well as sets out policy directions for Türkiye under its Artificial Intelligence Action Plan (2026–2030).