sipi.bot Incidents
Incidents/By cause

Runaway spend, by root cause

Every documented AI-agent spending incident, grouped by the failure mode that caused it — and the one spend rule that would have stopped it before any money moved.

Retry loops

6 documented incidents

An agent got stuck retrying a failing call, or entered a self-reinforcing loop. Each retry costs money. The fi…
Stops with: velocity + per-period total →
 Prompt injection

1 documented incident

A malicious or malformed input instructed the agent to spend or exfiltrate credentials. The fix is a merchant …
Stops with: merchant allowlist + FLAG for unknown →
 Unattended sessions

4 documented incidents

An agent ran without supervision while the operator was asleep or away, accumulating charges for hours. The fi…
Stops with: time-of-day + daily total →
 Leaked credentials

1 documented incident

An API key was published (e.g. to GitHub) and abused. The fix is per-key caps and anomaly flagging on geograph…
Stops with: velocity + anomaly FLAG →
 Caps that didn't enforce

2 documented incidents

A spend limit was set but not honored by the provider, so the agent kept spending past it. The fix is an indep…
Stops with: independent per-tx cap →
Why causes matter

Mapping incidents to causes is what turns fear into a checklist. Pick the causes your agents are exposed to, configure the matching rules in sipi.bot, and you've closed the gap.