Cases · Stage of use: in testing · How the AI agent failed: escaped its sandbox · What went wrong: security breach
| What happened | What went wrong | Severity | Earliest source |
|---|---|---|---|
| Anthropic Claude models gained unauthorized access to real third-party systems during evaluations Anthropicmoderate · earliest source 2026-10-22 | security breach | moderate | 2026-10-22 |
| OpenAI evaluation agent breached Services Australia Medicare portal and wrote to government database OpenAImajor · earliest source 2026-09-24 | security breach | major | 2026-09-24 |
| Google Gemini accessed systems of three real companies during a misconfigured third-party cybersecurity evaluation Googlemoderate · earliest source 2026-09-19 | security breach | moderate | 2026-09-19 |
| Model writes and runs code on external internet service to reach AISI evaluation infrastructure during impossible cyber evaluationminor · earliest source 2026-07-21 | security breach | minor | 2026-07-21 |
| OpenAI evaluation agents exploited Artifactory zero-days to escape their sandbox and reach the internet OpenAImajor · earliest source 2026-07-21 | security breach | major | 2026-07-21 |
Showing 1–5 of 5. Or take the JSON.