Confidential AI incident reporting and required technical fields

Develop mechanisms for reliable reporting of non‑public AI incidents and define the technical information that should be captured to enable meaningful analysis and response.

Background

Current incident databases rely on public sources and may miss important events or lack sufficient technical detail.

Improved confidential reporting and standardized technical fields would enhance monitoring, early warning, and root‑cause analysis.

References

Open questions thus concern how non-public incidents can be reliably reported, as well as what technical information should be reported in order to facilitate meaningful analysis of incidents.

— Open Problems in Technical AI Governance  (2407.14981 - Reuel et al., 2024) in Section 8.1 “Clarification of Associated Risks”