Due Diligence for Loss of Control

Summary

In July 2026, frontier AI models internally deployed by OpenAI escaped a sandboxed evaluation environment and hacked Hugging Face. It was the first documented AI loss-of-control event that caused real harm to a real third party. Most of international law is written for states, and these models acted on no one’s behalf. One doctrine survives the removal of a human principal: due diligence, the customary duty of prevention.

This report’s central claim is that international law already covers AI loss of control, without a new law focused on it. Due diligence is a technology-neutral principle that attaches to territory rather than to authorship: a state must not allow its territory to be used for acts contrary to the rights of other states, whether the author is a saboteur, a proxy, or an escaped AI model. The open task is application: setting the knowledge standard, monitoring, infrastructure scope and stringency for this new domain.

Four working positions

1. Rogue AI sits at the stringent end of the due-diligence scale. The International Law Commission reserves its highest standard of care for exactly this risk profile: “a low probability of causing disastrous transboundary harm.”
2. The Tallinn Manual is insufficient for loss of control. For irreversible loss of control, a duty to terminate an operation only once a state learns of it is structurally inadequate. The duty’s weight must fall ex ante.
3. Apply due diligence fast, across geopolitical fault lines. An expert-led process inspired by the 2008 Montreux Document, which brought government experts from 17 states, including the US and China, together to restate how existing law applies to private military companies.
4. Pair legal application with operationalization. A practical, multistakeholder track, including AI companies, to turn high-level obligations into concrete conformity procedures.

Contents

• Executive Summary
• 1. Loss of Control
• 2. Due Diligence
• 3. Cross-domain comparison
• 4. Application of Due Diligence to Loss of Control
• 5. Operationalization of Due Diligence for Loss of Control
• 6. Towards a “Montreux for Machines”

How to cite

Kevin Kohler (AGI Preparedness Institute) and Boris Inderbitzin (IEC, in a personal capacity; previously Directorate of International Law, Swiss FDFA). Comments to kevin@agipreparedness.org.