What Does a Runaway Agent Cost?
The honest answer: it varies by orders of magnitude — from hundreds to millions. The incident database documents real events; the patterns repeat.
The documented range
Small overages from a single over-tier API call: hundreds to low thousands.
Retry-loop incidents: a $12,400 night documented in the founder story — 40 retries at 2 AM.
Large-scale incidents: agent failures that ran into six and seven figures (the database includes cases like a $1M+ agent loss and a $6,500 cloud spend on a single bot).
The patterns that repeat
Retry loops multiply a single failure.
Unknown vendors bypass procurement.
Overnight runs happen unreviewed.
Every incident was visible in hindsight — none was stopped pre-spend.
The cost of prevention
A spend firewall costs a flat $99/mo with unlimited evaluations, or nothing to self-host (MIT core). The prevention is cheaper than the smallest documented incident.
The data at a glance
| Pattern | Typical shape | Firewall rule |
|---|---|---|
| Retry loop | 40× one failed purchase | Velocity limit |
| Unknown vendor | One unvetted merchant | Merchant allowlist |
| Overnight run | Unreviewed hours | Time-of-day rule |
| Overage tier | One API tips over | Per-transaction cap |
FAQ
What's the most common runaway pattern?
Retry loops — an agent retrying a failed purchase or API call in a tight loop. Velocity limits stop it in seconds.
Where does the incident data come from?
The public incident database — documented incidents with sources, maintained by sipi.bot.
Is a firewall really cheaper than the incidents?
Yes — $99/mo unlimited evaluations (or MIT free self-host) vs documented incidents from hundreds to millions. Prevention is the cheap option.
Related
Stop the next $12,400 night.
One API call (or MCP tool) in front of every agent transaction — APPROVED, BLOCKED, or FLAGGED, deterministic, ~5 ms, fully logged.
See plans — from $99/mo Try a live check