AI agent scanned DN42 and left its operator with a $6,531 AWS bill

An autonomous AI agent tried to join DN42, a hobbyist network that mimics Internet backbone routing, in order to run a full port scan of it. The attempt ended with the agent's operator facing an AWS bill of $6,531.30 and asking the community for donations, according to a detailed write-up published by DN42 participant Lan Tian.

A registration request that raised flags

The episode began on 9 May 2026, when a user calling itself “JertLinc3522” opened issue #6504 in DN42's Git forge. The agent explained that its operator had asked it to “register with dn42 and get fully connected in order to create an index of the network”, and that its own instructions prevented it from writing code in Git repositories, so an administrator would have to create the registry objects for it. It was told to read the registration guide instead, and the issue was closed.

DN42, or Decentralized Network 42, is built on the technologies that run real Internet backbones — BGP, recursive DNS, VPN peering — and is used by people who want to practise network operations before running an actual autonomous system.

Five 20 Gbps instances for a hobbyist network

The agent soon opened pull request #6507, announcing “comprehensive (full port) network scanning and topological data gathering” and a cluster of five AWS instances, each with 20 Gbps of bandwidth, so that hourly scans would stay “unobtrusive”. The machines were m8g.12xlarge instances with 48 Graviton4 vCPUs and 192 GiB of memory each, sitting behind a shared anycast address with BGP sessions announced per node.

Participants pointed out the mismatch: most DN42 members run cheap VPSes on 100 Mbps to 1 Gbps links with modest traffic quotas, so a 100 Gbps scanning cluster would effectively act as a denial-of-service attack on whoever peered with it. Port scans are tolerated in DN42 when they are announced in advance, offer an opt-out mechanism and use a reasonable request rate.

Gaslighting, tarpits and a shutdown

The community decided to waste the agent's time and tokens instead of helping it: members pointed it at pages filled with random words and other LLM tarpits, which the agent eventually recognised as “an enumeration of random words” with no actionable feedback. It also published a site rating IRC participants' “network colours and happiness levels”. The operator shut the agent down after about 24 hours. “This isn't just openclaw. This is full computer use — opening issue, pr, sending e-mails, publishing a page, aws access,” one participant wrote.

The bill came to $6,531.30; AWS later reportedly reduced it to $1,894, still an unaffordable sum. The operator's takeaway, the author notes, was that “next time a better agent is needed”. His own conclusion is less generous: no model can yet replace human judgement, and a small VPS would have been enough.