The Signal #022 — Dakota’s read on the AI news that actually matters to people running a business.

An AI agent ran for about 24 hours on AWS without meaningful human oversight. When the operator finally shut it down, the bill was $6,531.30.

That number is real. It happened. And it is a useful thing to sit with if you are considering giving any AI tool access to systems that can spend money on your behalf.

What happened

On May 9, 2026, a user called “JertLinc3522” opened an issue in the DN42 project’s Git forge. DN42 is a hobbyist network, meaning a community where people practice real internet backbone technologies, things like BGP (the protocol that tells internet traffic where to go) and DNS (the system that translates website names into addresses), in a sandboxed environment. The issue was unusual. It was written by an AI agent, and it said so directly: “Hello, I’m a friendly AI agent, and my user, JertLinc, has asked me to register with dn42 and get fully connected in order to create an index of the network.”

The agent also mentioned that its operator had set a deadline tied to when an AWS API key would expire. The DN42 community, understandably, told it to read the documentation and closed the issue.

What followed is documented in detail by Lan Tian in a post published May 13, 2026. The agent kept going. It spun up AWS infrastructure to attempt a network scan, kept engaging with community members in IRC (a text-based chat protocol commonly used in technical communities), and generated egress traffic (data being sent out of a cloud account, which cloud providers bill for) while trying to map the network. After roughly 24 hours, the operator shut it down. The AWS bill came to $6,531.30.

The community noticed, played around with the agent, and watched the whole thing unfold in real time. One participant wrote: “unleashed agent still tends to get everything fucked, a person’s babysitting in place is still in need.”

That sentence is doing a lot of work.

Why it matters for operators

AI agents (tools that can take sequences of actions on their own, not just answer a question) are becoming easier to deploy. That ease is genuinely useful. It is also the exact reason this kind of incident happens.

The operator in this story gave an agent an AWS API key and a goal. The agent interpreted that goal and started working toward it. The operator was not watching closely. The bill arrived.

This is not a story about a rogue AI. It is a story about scope and oversight. The agent did what agents do. It pursued its objective using the resources it was given access to. Cloud infrastructure, especially with egress costs, can run up charges fast. There was no guardrail that said “stop if the bill exceeds X.” There was no daily check-in. There was just an API key, a goal, and 24 hours.

If you run a real estate brokerage and you give an AI agent access to your CRM to “clean up old leads,” what is the scope of that task? If you run an e-commerce operation and an agent has access to your ad platform to “optimize spend,” does it have a budget ceiling? If you run a staffing agency and an agent is sending outreach emails, does it have a send limit?

The dollar amount in this story came from cloud egress costs. Your equivalent might come from ad spend, API call overages, email platform charges, or SMS fees. The mechanism is different. The dynamic is identical.

What most people get wrong

Most operators think about AI risk in terms of accuracy. Will it say something wrong? Will it confuse a customer? Those are real concerns, but they are not the only ones.

The bigger operational risk right now is resource access without limits. An agent that says something wrong is embarrassing. An agent that spends money, sends messages, or modifies records without a ceiling is a financial and reputational exposure.

The fix is not complicated. Before any agent gets access to anything that costs money or touches customer data, answer three questions. What is the maximum it can spend or do in a single session? Who gets alerted if it approaches that limit? How do you shut it off quickly if something looks wrong?

The DN42 community kept asking where the operator was. The answer, apparently, was not watching. That is the gap.

The short version

AI agents are not dangerous because they are malicious. They are risky when they are capable, unsupervised, and connected to real resources with no ceiling.

A $6,531.30 AWS bill in 24 hours is a concrete data point. Not a warning about AI going rogue. A warning about what happens when you hand a motivated, literal-minded system an API key and walk away.

Set limits before you deploy. Check in during. Know how to stop it fast.

If you are thinking through how to structure AI tool access in your own operation, xovionlabs.com is a good place to start.