A Common Standard for AI Incident Accountability

A Common Standard for AI Incident Accountability Abstract This technical-policy paper proposes a cross-institutional framework for the reporting, disclosure, review, and interpretation of material incidents involving advanced artificial intelligence systems. The framework is designed for use across research organizations, AI developers, evaluators, cloud and infrastructure providers, cybersecurity practitioners, public-interest institutions, standards bodies, and oversight entities. Its purpose is to support consistent incident documentation while balancing transparency, technical rigor, public accountability, and responsible handling of security-sensitive information. The paper defines criteria for determining when an AI-related event should be considered reportable, introduces a four-class severity model, specifies minimum incident-report contents, and outlines evidence-preservation requirements covering model metadata, prompts, tool activity, security telemetry, infrastructure configuration, human intervention, and chain-of-custody information. Particular emphasis is placed on distinguishing the behavior of a model from the behavior of the larger deployed system in which it operates. Accordingly, the proposed reporting framework treats model identity, agent scaffolding, runtime configuration, permissions, tools, credentials, monitoring systems, and operational context as separate but related elements of incident analysis. The paper also proposes a two-layer disclosure model consisting of: 1. a public incident report intended to provide meaningful transparency and accountability; and2. a restricted technical annex for information whose immediate public disclosure could create additional security, privacy, or operational risk. To support reproducibility and consistent interpretation, the framework includes guidance for reading AI incident reports, separating observed actions from inference, identifying authorized operational boundaries, evaluating failed and successful controls, assessing uncertainty, and interpreting remediation commitments. The final sections propose principles for cross-institutional governance and provide a reusable public incident-report template suitable for adaptation by research organizations, AI laboratories, evaluators, infrastructure providers, and other entities operating advanced AI systems. Purpose The objective of this work is not to establish a binding regulatory or technical standard. Rather, it is to provide a structured starting point for public discussion, technical review, institutional experimentation, and future standards development. The proposed framework is intended to make AI incident reports more: - comparable across organizations;- technically meaningful;- transparent to affected parties and the public;- useful to independent investigators;- compatible with responsible vulnerability disclosure; and- suitable for longitudinal analysis across multiple incidents. Scope This paper focuses primarily on incidents involving advanced AI systems operating in evaluation, research, staging, production, or agentic execution environments. Examples may include: - material violations of intended operational boundaries;- unauthorized access to external systems;- defeat or circumvention of monitoring or containment controls;- unexpected acquisition or use of permissions or credentials;- significant vulnerability discovery or exploitation;- material effects on third parties;- security-relevant failures involving autonomous or semi-autonomous agents; and- other events whose recurrence could create meaningful safety, security, or public-interest concerns. Routine model errors, harmless evaluation anomalies, and contained experimental failures are not automatically treated as public incidents, although they may remain appropriate for internal evaluation records. Proposed Contributions This work introduces: - a definition of reportable AI incidents;- a four-tier severity and disclosure classification;- a minimum incident-report schema;- model and system identification requirements;- evidence and metadata preservation guidance;- a framework for separating public disclosure from restricted technical evidence;- guidance for interpreting AI incident reports;- a reusable public incident-report template; and- principles for cross-institutional governance and review. Status This document is a proposed technical-policy framework for research and public discussion. It is not an official standard, regulation, government document, industry requirement, or endorsed policy of CERN, Zenodo, OpenAI, or any other institution. The framework is intended to be revised through criticism, implementation experience, technical review, and cross-disciplinary discussion. Version Version 1.0 Initial public release.