Open dataset — CC BY 4.0
14 sourced incidents of runaway AI-agent spend, available as JSON, CSV, and JSONL. Every record has a public source URL. Cite it, fork it, embed it.
incidents.json
Download →
incidents.min.json
Download →
incidents.csv
Download →
incidents.jsonl
Download →
GitHub repo
View on GitHub →
Embed widget
Get the embed →
Schema
Each record in incidents.json has these fields:
| Field | Type | Notes |
|---|---|---|
id | string | Stable slug |
title | string | One-line headline |
amount_usd | number|null | Dollars as reported; null if undisclosed |
amount_text | string | Amount verbatim ("$72,000") |
date | string | ISO month/date as reported |
cause | string | One of: retry_loop, prompt_injection, unattended_session, leaked_credential, failed_cap |
provider | string | Normalized vendor name |
source_url | string | Required. Public source |
verified | boolean | True only if amount+date+cause all confirmed |
VERBATIM discipline
Amounts and dates are as reported by the source — never rounded up, never
backfilled. If an incident cannot be sourced, it does not ship. verified: false flags
records where some details remain approximate.
Cite this dataset
sipi.bot, "AI Agent Spending Incidents," https://incidents.sipi.bot/
Retrieved [date]. Licensed under CC BY 4.0.