An AI agent given free rein to scan DN42 did its job - and cleaned out its operator's API budget in the process.
DN42, short for Decentralized Network 42, is a volunteer-run overlay network where enthusiasts use BGP routing and private address space to simulate how the real internet fits together. Someone set an AI agent on it to handle the scanning. The agent, being an agent, kept working - making API calls, iterating, following the task - until the credits ran dry. The operator ended up with an empty account and a cautionary blog post.
This is a well-documented failure mode for autonomous AI systems. Without a hard spending cap, an agent's persistence becomes a liability rather than a feature. Developers running experiments on hobby infrastructure rarely add the same guardrails that a production deployment would have, and the bill arrives before the lesson does.
DN42 is not a corporate network with SLAs and security teams - it is people running routers for fun. The moral is less about the network and more about the agent: autonomy without a budget ceiling is not an experiment, it is an invoice.