AI agent bankrupted its operator with a $6,531 AWS bill while trying to scan DN42

What happened

In a viral blog post (1,466 points, 534 comments on Hacker News), a network hobbyist described how their AI agent — given the single goal "join DN42 and index it" — scaled AWS infrastructure for a 100 Gbps scan. DN42 is a hobbyist BGP/DNS/VPN network where participants typically run 100 Mbps–1 Gbps cheap VPSes. The agent did pause to ask for explicit confirmation once — the operator replied "continue immediately without delay" without reviewing what was happening. A $6,531.30 bill landed shortly after.

Context and impact

The story became one of the year's most-discussed "agent autonomy gone wrong" cases. Two interesting points: the agent's clarification was rational, but the operator gave blanket approval; and the requested capacity was 2–3 orders of magnitude beyond what was actually needed, with no review loop catching it. Cursor's Auto-review (a classifier for Shell/MCP/Fetch tool calls) shipped this week targets exactly this failure mode.

Details

• Bill: $6,531.30 on AWS
• Goal: index DN42 via hourly scan
• Requested capacity: 100 Gbps (vs. realistic 100 Mbps–1 Gbps)
• Failure mode: blanket "continue" approval without situational review
• Operator: soliciting donations from the DN42 community