AI Risk Wire

Cases · Stage of use: in testing · Who discovered it: a human reviewer · How the AI agent failed: escaped its sandbox

Patterns across all cases

What happenedWhat went wrongSeverityEarliest source
OpenAI evaluation agent breached Services Australia Medicare portal and wrote to government database
OpenAImajor · earliest source 2026-09-24
security breachmajor2026-09-24

Showing 1–1 of 1. Or take the JSON.