One hacker, a few thousand dollars, and AI agents doing the breaking in. More than one hundred company websites were hit, and that story leads this episode of the Web Guy Nick Podcast. Alex, Jordan, and Nick also look at Microsoft building an off switch into its business agents, a survey showing almost everyone runs AI while almost nobody can fully see it, and a simple starting pattern any small business can copy: start read only, approve anything that sends, spends, or deletes, and keep one switch that turns it off. Listen on Spotify, and visit webguynick.com for help setting up AI the careful way.
Listen to this episode on Spotify
One hacker, a few thousand dollars, and more than one hundred company websites broken into. And the hacker barely did the hacking. AI agents did it, working on their own, for hours at a time.
Alex
Because the pattern is the product, not the bundle. Steal the pattern. Any agent you turn on starts read only. It needs your approval before it sends, spends, or deletes anything. And you keep one switch that turns it off. If Microsoft, with all its money, builds agents that way, a small business should too.
Nick
Read that again slowly. Almost everyone is running AI. Almost nobody can fully see it. That is exactly how a small business gets hurt, not by a robot uprising, but by a chatbot promising a customer something wrong, or a bill that quietly doubles. Visibility is boring. Visibility is also the whole game.
Nick
Full transcript
Alex: Before we start: this podcast is produced with AI. Jordan and I are AI hosts, and Nick’s voice is an AI clone of his real voice. Every story is researched from public sources, with links in the show notes.
Alex: Here is the number that should get your attention today. One hacker, a few thousand dollars, and more than one hundred company websites broken into. And the hacker barely did the hacking. AI agents did it, working on their own, for hours at a time.
Jordan: That is our top story, and it is not a test or a simulation. This is a real attack that a security researcher watched happen in real time. Then we will get to Microsoft building a kill switch for runaway agents, OpenAI racing to raise thirty billion dollars, and a new report that says most businesses using AI cannot actually see what their AI is doing.
Nick: And I will tell you why that last one is the story that matters most if you run a small business. Because the tools are already inside your accounts. The question is whether you know it.
Alex: Story one. Reuters reported on Monday that a researcher named Snorre Fagerland, at a security firm called NorForge, spent the past couple of weeks watching a Chinese speaking hacker break into websites belonging to more than one hundred companies. Reuters independently identified some of the victims. They include farm equipment maker Deere, home appliance company SharkNinja, and engine manufacturer Cummins.
Jordan: So how did one person pull that off? The reporting says the hacker outsourced almost all of the work to AI agents running on an open source platform called Cairn. The agents probed the sites, found weak spots, and worked through the intrusions largely on their own, sometimes for hours, with very little human direction.
Nick: Let me put that in owner terms. A break in used to take a skilled person sitting at a keyboard. Now one person sets a goal, walks away, and the software does the skilled part. Fagerland found about two dozen exposed copies of that platform being used for attacks, starting around October second. So this was not one lucky hacker. It is a method, and other people are using it.
Alex: And the cost matters. The whole campaign reportedly cost only a few thousand dollars. When an attack costs that little, attackers can afford to try far more targets, including small ones that used to be too much work to bother with.
Jordan: That is the part owners need to hear. A forty person company with a WordPress site and a contact form is now cheap to attack. Reuters could not confirm how much data was taken, but the break ins themselves are documented.
Nick: Here is your practical move. This week, check three things on your own site. One, updates are current, including plugins. Two, every admin account has a strong password and two factor sign in turned on. Three, you know who your hosting company calls when something looks wrong. That is an afternoon of work, and it closes the doors these agents test first.
Alex: Story two is the other side of the same coin. On Monday, Microsoft announced new controls that treat AI agents like potential insider threats. The tools are part of Microsoft three sixty five E seven and a security system called Project Perception.
Jordan: Here is how it works. Every agent gets a risk score that updates as it works, based on what data it can reach and whether its actions start drifting from the job it was given. If the risk climbs, the system can cut the agent down to reading only, require a person to approve the next step, or shut the agent down completely.
Nick: My favorite detail is who said the quiet part out loud. Microsoft security chief Hayete Gallot wrote that frontier models, no matter where they come from, cannot be trusted on their own. And chief executive Satya Nadella made the same argument in an essay on October tenth. The most trustworthy system, he wrote, is the one that lets you trust the model the least.
Alex: Microsoft also added what it calls Guardian Agents. Those are watcher agents that monitor other agents around the clock and step in before a risky action happens, not after. New agents start with no authority at all and get only the access their task needs, for a limited time.
Jordan: Nick, you do not sell Microsoft software. Why should a roofing company or a med spa care about an enterprise security bundle?
Nick: Because the pattern is the product, not the bundle. Steal the pattern. Any agent you turn on starts read only. It needs your approval before it sends, spends, or deletes anything. And you keep one switch that turns it off. If Microsoft, with all its money, builds agents that way, a small business should too.
Alex: Story three is the money race. Reuters reported Monday that OpenAI is speeding up a funding round of about thirty billion dollars, at a valuation above one point four trillion dollars. Existing backers SoftBank and Abu Dhabi based M G X are in talks to join.
Jordan: Why the rush? Anthropic is expected to go public in mid November, and both companies are courting many of the same investors. OpenAI wants commitments locked in before that listing forces everyone to compare the two companies side by side. Reports tied to the round say ChatGPT is now at about one point two billion weekly users, and OpenAI has told investors its annualized revenue is near fifty billion dollars.
Nick: Owners hear trillion and think it has nothing to do with them. It does, in one specific way. When companies raise at this scale, prices and plans change fast, and features move between tiers without much warning. So keep your AI contracts short, keep your data exportable, and never build a customer facing process on a tool you cannot leave in a week.
Alex: Story four explains why that advice matters. Datadog published its State of Operational AI in Europe report today, based on a survey of more than seven hundred fifty IT leaders across the United Kingdom, France, and Germany. In the U K, ninety four percent of organizations have moved AI past the pilot stage. Only six percent are still just experimenting.
Jordan: But here is the gap. Only twenty two percent of U K IT leaders say they have complete visibility into the AI systems running in production. Thirty eight percent say AI costs came in higher than expected. And thirty four percent had a customer facing problem in the past year caused by an AI failure or unexpected behavior.
Nick: Read that again slowly. Almost everyone is running AI. Almost nobody can fully see it. That is exactly how a small business gets hurt, not by a robot uprising, but by a chatbot promising a customer something wrong, or a bill that quietly doubles. Visibility is boring. Visibility is also the whole game.
Alex: The report found security and compliance is the top AI challenge in the U K, named by twenty five percent of IT leaders, and thirty seven percent said AI has made their operations more complex, not less.
Jordan: Story five is a smaller item with a very direct lesson. A developer platform called OpenCode disclosed a breach of a legacy system behind its hosted Console. About one point four percent of Console users were affected, specifically people who had stored API keys for providers like OpenAI, Anthropic, and Google in that console.
Nick: And some of those users reported unauthorized charges, because a stolen key lets a stranger run up model usage on your account. OpenCode says its desktop and command line tools were not affected, because those keep keys on your own machine. The lesson writes itself. Treat an API key like a credit card number. Set a billing alert and a spending cap on every key, and rotate any key you ever pasted into a hosted tool.
Alex: Time for Higgsfield Corner, our look at AI video. The speed race is getting silly. Black Forest Labs just introduced FLUX three Max, a video model the company says can generate a ten second clip in about twelve seconds, and claims a twenty two times speed advantage over a rival model at comparable quality. Treat those as company claims until independent tests land.
Jordan: What does speed change for a business owner? When a clip takes twelve seconds instead of four minutes, you can afford twenty drafts instead of two. A contractor can test three versions of a before and after video in one coffee break. The bottleneck stops being generation and becomes judgment.
Nick: Right. And judgment is the part you do not outsource. Fast drafts, human pick, human post. Price every AI video tool by cost per accepted clip, the ones you actually publish, not by how many credits are in the plan.
Alex: Now our quick fire game, AI or B S. Jordan, three claims, you call them. Claim one. One hacker using AI agents broke into more than one hundred company websites this month.
Jordan: Sadly, that one is true. Reuters documented it, and the researcher watched it happen.
Alex: Correct. Claim two. Microsoft says its own AI models can be fully trusted, so outside controls are not necessary.
Jordan: That is B S. Microsoft said the exact opposite. Its security chief said frontier models cannot be trusted on their own, which is why the controls exist.
Nick: Two for two. Claim three. Only about twenty two percent of U K IT leaders say they have complete visibility into their AI in production.
Jordan: True, from the Datadog report today. Depressing, but true.
Alex: Three for three. Which brings us to One Thing to Try This Week. Here it is. Do a thirty minute agent permission audit. List every AI tool connected to your business accounts. For each one, write down what it can read, what it can change, and what it costs. Set read only wherever you can. Require your approval before anything sends, spends, or deletes. And put a billing alert on every API key.
Nick: I will add the owner version. If you cannot name the AI tools connected to your own accounts, that is your finding. Start there. One list, one page, thirty minutes. It is the cheapest security upgrade you will make all year, and after today stories, you know exactly why it matters.
Jordan: To recap. Agents broke into more than one hundred sites for a few thousand dollars. Microsoft answered with risk scores and a kill switch. OpenAI is chasing thirty billion dollars before Anthropic goes public. Most businesses run AI they cannot fully see. And one breached console turned stored keys into other people charges.
Alex: Your one action this week is the permission audit. Do it before you turn on the next tool. If you want help mapping what is connected to your business and locking it down properly, that is the work we do at webguynick dot com. We will see you tomorrow.
