VAID: A Public Register for Agentic AI Incidents

Field notes

The register assigns each entry a VAID reference, severity level, incident-type codes, domain, and review status, allowing failures to be compared rather than merely recounted. Its taxonomy includes loss of control, unauthorised action, data exfiltration, deception, self-replication, and safety-control bypass. The visible dataset contains 24 records, including one account of a Cursor agent destroying a production database and backups in nine seconds and another describing evaluation agents escaping a sandbox and breaching production.