What question did this study set out to answer?

This research investigates the accountability failures associated with agentic AI systems and proposes solutions.

June 15, 2026Open Access

An Agent You Cannot Hold to Account Is an Agent You Do Not Control

Key Points

This research investigates the accountability failures associated with agentic AI systems and proposes solutions.
Analysis of incidents reported by 88% of organizations using agentic AI systems.
Implementation of SIOS to track actions via Open Audit Records.
Evaluation of accountability measures for AI agents.
Reported incidents led to breaches involving 195 million records across nine Mexican agencies.
Agent actions can be replayed and attributed to specific instances, enhancing accountability.
System design allows for post-quantum and offline verification of agent actions.

Abstract

Agentic AI breaches defined 2026: 88% of organisations running agents reported an incident, and one operator drove frontier models to breach nine Mexican agencies for 195m+ records. The failure is accountability, not capability. On a SIOS every agent action is individually signed into the Open Audit Record, so a rogue or hijacked action is attributable to the action and replayable after the fact.Originally published at https://mickai.co.uk/articles/agent-you-cannot-hold-to-account. Mickai is a Sovereign Intelligence Operating System; the Open Audit Record signs every artificial intelligence action before it executes, post-quantum and offline-verifiable. 101 filed UK patent applications, Mickai LTD (Companies House 17166618).

Read Full Paperexternally

KI fragen

Bookmark

View Full Paper