How Should a Small Business Prepare for Attacks That Now Move at AI Speed?

Attackers now use AI to move in hours, not weeks. Here is what a Northeast Indiana owner can inventory, approve, and patch before the next warning lands.

How Should a Small Business Prepare for Attacks That Now Move at AI Speed?

A small business prepares for AI-speed attacks by knowing where its access keys live, setting rules for what automated tools may do, and holding a fixed patch window. That is the practical lesson from the October 22 episode of The AI Daily Brief, which describes a flaw in PHP, a widely used web programming language, that could affect as many as 70% of websites and web apps.

Knowledge of the flaw spread roughly ten hours before public disclosure, which left most owners with little time to react. The host frames the moment as a shift from manageable exploits to a race in which AI attacks and AI defenses coordinate at inhuman speed. For an owner in Northeast Indiana, the useful question is what the business controls before the next warning reaches the headlines.

Key Takeaways

  • Attack speed has changed, and the episode describes exploit attempts counted in the millions within hours of disclosure.
  • Early knowledge of a flaw can spread before a patch reaches most servers, so waiting for the public announcement puts an owner behind.
  • Keys, tokens, and environment variables are the prize, and the Railway and OpenCode incidents both pointed directly at them.
  • A patch window planned in advance matters more than a fast reaction improvised on the day an alert goes public.
  • Approval rules for automated tools limit damage when something goes wrong and stop bad actions before they reach customers.
  • An AI teammate can watch, draft, and prepare, while a person approves every change that touches a live system or a customer.

Why do attacks now move faster than a small business can react?

Attacks now move faster than a small business can react because AI tools let attackers scan, test, and exploit systems at a scale that no small team can match by hand. The episode describes a sequence that unfolded over hours rather than weeks. Cloudflare, as reported by its chief executive, saw thousands of attempts to exploit the flaw in the ten hours before the announcement, tens of thousands in the 45 minutes after, hundreds of thousands within two hours, and then millions.

Those figures originate from one network, so they are not a census of the whole internet, but they still demonstrate the pace.

The timing was also unusual. According to the episode, a researcher at Caliph and, independently, the security firm Wiz found the flaw about two months earlier, using GPT 5.6 Sol. OpenAI then spent those two months coordinating with PHP's maintainers, Wiz, and Cloudflare so that mitigations and a patch would be ready before disclosure.

That coordination worked in one sense, because a fix existed at announcement. Yet the host notes that knowledge of the exploit spread roughly ten hours ahead of disclosure, before most servers could be updated. An owner who waits for the public announcement therefore starts behind everyone who heard early.

The lesson for an owner is not that every warning is a catastrophe. The host is openly and reasonably skeptical of sky-is-falling cyber stories, and says this one drew alarm from people who know the field well. The lesson is that the gap between a flaw existing and attackers using it has shrunk considerably. Preparation has to happen before the headline, because the headline now serves as the starting signal for attackers as well as defenders.

What does fleets versus swarms mean for an owner?

Fleets versus swarms describes defenders running a governed group of AI systems while attackers run cheap, persistent swarms that never stop. Tal Hoffman of Enclave framed the contest that way, and the phrase identifies something real. Defenders have structure, permissions, and review, while attackers have volume and patience. The episode quoted Hoffman's conclusion, "Whoever coordinates faster wins," and that line is the most useful takeaway for a business that cannot build a fleet of its own.

A small business can still decide who coordinates its response. For a team of five to fifty people, coordination comes down to three things: a concise list of systems, one named person who receives alerts, and a written sequence for the first hour after a warning. Without those three pieces, even a capable tool produces noise, and noise slows the response at exactly the moment when minutes count. Coordination is primarily a matter of ownership and habit, which costs far less than a dedicated security budget.

The episode also raised a structural point that is worth keeping in view. One commentator argued that markets may not yet price in how quickly banks and healthcare organizations will need to merge to gain enough scale to defend against AI-driven vulnerabilities. The implication for a small firm is narrower but similar in spirit. Scale is no longer the only defense, and clear ownership of security tasks, combined with a weekly routine that runs whether or not anyone feels worried, now matters as much as the size of the budget.

Which keys and accounts do attackers want most?

Attackers most desire the credentials that open other systems, including API keys, database logins, and secret values stored on hosting platforms. Two incidents in the episode show why. A recent attack on OpenCode reached about 1.5% of its users, who had their API keys stolen and then saw large bills run up on their accounts.

Days later, Railway, a hosting service similar to Vercel, said an unauthorized party may have accessed encrypted environment variables and decrypted some of them. Those variables can hold API keys and database or registry credentials that open the doors to the rest of a business.

Railway told its customers to rotate their API keys and roll every environment variable stored on the platform. The episode also quoted an OpenCode engineer who warned that attackers will target every surface and every employee, and that projects forgotten for years can be found and used against their owners.

That warning points to a straightforward exercise any owner can run this month. List every service that holds a login, a key, or a payment token, and record who created each one, when it last changed, and whether it still serves a purpose.

Forgotten accounts deserve special attention, because they are the easiest to overlook. An old website plugin, a retired contractor login, or a test database can become the quickest route into a business that has since moved on. Rotating keys is rapid when the list exists and slow when it does not.

An owner with a current list can respond in minutes, which is the difference the episode's entrepreneurs described when they rotated every key across many projects at once. The same list also shows which tools are connected to customer data, which is the first question any incident response must answer.

Why can the patch itself be dangerous, and what should owners do first?

The patch can be dangerous because a fix often reveals the flaw to anyone who reads it closely, so maintainers time releases with care. The episode explained that the patch makes the exploit obvious, and releasing it too early would hand a flood of unskilled attackers a map to every unpatched site.

That reasoning is why the NSA advised organizations to take systems offline until a fix landed, and why the FBI's Cyber Division asked for reports of intrusions. The PHP2Shell timeline for site owners walks through the patch details for anyone running PHP.

A small business is rarely in a position to make that call for its vendors, but it can prepare to act the moment a fix becomes public. Begin with an inventory of every public website, booking page, and customer portal, and record the software and version each one runs.

Ask each vendor, in writing, how and when it applies security updates and who receives the notice. Then establish a patch window, such as the first morning after an alert, when someone can apply the update and confirm that the site still works for customers.

Owners should also plan for a slow vendor, because slowness is common even among capable companies. One technology employee quoted in the episode replied to a warning that he would patch in a few days because he was out of office. That answer is not safe for a business that depends on a public website for leads or payments. A plan that names a backup person, a contact at the hosting company, and a way to take a site offline briefly prevents one person's absence from becoming a breach.

What else was in the October 22 episode?

The episode opened with a short set of headlines before the main cyber story, and several of them connect to the same theme of AI systems acting on people's behalf. Anthropic announced Claude Duet, a personal AI system aimed at professionals and small businesses. It sends a morning brief across connected apps, interrupts only when something needs attention, and is available through a waitlisted research preview. Many listeners noticed that it runs on Anthropic's Opus 5.5 model, which the host considered the real differentiator.

Walmart announced that it will allow customers link personal AI systems to their accounts with spending limits, starting next month. Its terms require purchases to be approved by default, forbid actions beyond what the shopper intended, and bar changes to payment methods, which makes them a useful template for any owner who lets software touch money. Meta's earnings call added a data point that surprised many e-commerce veterans: about a quarter of active users of its Muse assistant use it for shopping, with average purchases of about $200.

Meta also announced a new class of non-voting stock to help fund a data-center budget that now runs between $135 billion and $145 billion this year. The host read the move as a way to raise money without diluting the chief executive's voting control, and noted that Meta's stock rose overnight anyway. None of these headlines changes a small business's next step on its own. Together, however, they show how quickly AI systems are gaining authority over everyday work, and why approval rules matter more each month.

What does an AI teammate do with this, and what does it never do?

An AI teammate can watch for the signals this episode describes, draft the response, and organize the checklist, while the owner approves anything that changes a live system or reaches a customer. At Hey Button, the default is draft-and-approve, so nothing is sent without approval. In an illustrative example, read access lets the teammate review the systems inventory and the vendor list. Draft access lets it prepare a patch notice, a key-rotation checklist for the owner, or a reply to a customer who asks whether their information is safe.

The boundary matters as much as the capability itself. A teammate with read-only access can report that a key appears on an old page, but it cannot rotate that key on its own. A teammate with draft-and-approve access can prepare the rotation steps and then wait for the owner's yes before anything changes. Full access represents a separate decision, and most owners should refrain from granting it for security changes until they have watched the teammate handle real examples.

Hey Button's setup starts with a conversation about how work comes in, which tools you use, and where busywork piles up, and it sets approval rules before anything touches a customer. The reasoning behind approval-first setup explains why that order matters.

A teammate is not a substitute for a security professional. It cannot identify every vulnerability, and it can make mistakes, which is precisely why the approval step exists. The owner carefully reviews each proposed change, checks it against the inventory, and makes the decision. The teammate reduces the time spent gathering facts, and that time is often where small teams lose the race to an attacker.

Can a small business use AI to defend itself without giving up control?

A small business can use AI to defend itself without giving up control by limiting what the AI may do, keeping a written record of its decisions, and reviewing its work on a set schedule. The defensive value that the episode describes is real. One entrepreneur said that rotating every key across a large number of projects in minutes would have been a nightmare not long ago.

The host made a related point: a nuclear weapon cannot be stopped with another nuclear weapon, but an AI attack can be met with AI.

Control comes from the written rules, not from the tool itself. Write down the systems the teammate may read, the actions it may draft, and the actions that always return to you first. Reevaluate that list each quarter, because both the tools and the threats change quickly. Maintain the vendor list, and require every vendor to say plainly what its AI tools can access. Vendors that answer plainly are considerably easier to trust than vendors that answer with unqualified confidence alone.

Hey Button works this way, and the first step is a short questionnaire about what you would take off your plate first. If you want an AI teammate for your Northeast Indiana business that stays under your rules whichever model is behind it, start the Hey Button questionnaire and tell us which security task to tackle first. Setup and ongoing support are quoted around your workflows and integrations, and any software subscriptions or usage charges are identified before you commit.

Sources

  • The AI Daily Brief, "The AI vs AI Battle for the Internet," October 22, 2026 (episode and show notes): https://aidailybrief.ai/e/2026-10-22
  • OpenAI's blog post on the PHP2Shell vulnerability, as quoted in the episode.
  • Railway's disclosure post on the environment-variable incident, as quoted in the episode.
  • Greg Brockman, "Towards Securing the Internet," OpenAI, as quoted in the episode.
  • Matthew Prince, Cloudflare, on exploit attempt counts, as quoted in the episode.

Frequently Asked Questions

Usually only briefly, and only if the owner cannot confirm the software is patched or cannot patch quickly. A short, planned outage with a plain customer notice is often wiser than running an unpatched site for days. Decide that rule before any alert arrives, so the choice is not made in a panic.
Review it every quarter, and again within a day of any vendor breach notice. Remove accounts nobody uses, rotate keys that have not changed in a year, and confirm which people still need access. A written list turns the review into a routine task instead of an emergency search.
No. Read-only access lets it inventory systems and flag stale keys. Draft-and-approve access lets it prepare patch notices and rotation checklists for your sign-off. Full access is a separate decision, best made after you have watched the teammate handle real examples on your own systems.
A clear answer on which AI models it uses, where data is processed, what those tools can reach, and who approves changes. Ask for it in writing. A vendor that cannot describe its approval rules may not have them, while a specific, plain answer is a good sign.
Not always. A patch can reveal the flaw to attackers, so timing matters as much as speed. The safest plan keeps an update ready to apply, tests it when possible, and applies it in a window you choose. Speed matters most once an alert is public, and preparation matters before it.
Should a small business take its website offline during a serious warning?
Usually only briefly, and only if the owner cannot confirm the software is patched or cannot patch quickly. A short, planned outage with a plain customer notice is often wiser than running an unpatched site for days. Decide that rule before any alert arrives, so the choice is not made in a panic.
How often should a small business review its list of keys and accounts?
Review it every quarter, and again within a day of any vendor breach notice. Remove accounts nobody uses, rotate keys that have not changed in a year, and confirm which people still need access. A written list turns the review into a routine task instead of an emergency search.
Does an AI teammate need full access to help with security?
No. Read-only access lets it inventory systems and flag stale keys. Draft-and-approve access lets it prepare patch notices and rotation checklists for your sign-off. Full access is a separate decision, best made after you have watched the teammate handle real examples on your own systems.
What should a vendor be able to tell you about its AI tools?
A clear answer on which AI models it uses, where data is processed, what those tools can reach, and who approves changes. Ask for it in writing. A vendor that cannot describe its approval rules may not have them, while a specific, plain answer is a good sign.
Is a faster patch always the safest choice?
Not always. A patch can reveal the flaw to attackers, so timing matters as much as speed. The safest plan keeps an update ready to apply, tests it when possible, and applies it in a window you choose. Speed matters most once an alert is public, and preparation matters before it.

Hey Button

Want a teammate that handles this for you?

Answer a few questions about your business and Hey Button will show you what an AI teammate would take off your plate, with you approving anything that goes out.

Start the questionnaireSee what teammates do
Lucas M. Button

Written by Lucas M. Button

Founder, Hey Button

Lucas builds AI teammates for small businesses across Northeast Indiana and writes about what works, what doesn't, and what to hand off first. More about Lucas