A month ago we told CEOs that the flood of newly discovered software flaws was good news for their risk, at least for these early days. The reasoning was that AI was helping defenders find old flaws faster than attackers could use them, and the forecasting team at FIRST, the global association of security response teams, added a warning we passed along: the advantage might not last long, so seize it. It did not last long, but at least it did enable the largest “patch Tuesday” in history from Microsoft mid-October, a gleaming new layer of armor right at the moment the battle escalates.
In the space of five days this month, the FBI, Google, and the insurance broker Marsh each described the same shift from different angles. Attackers are now using AI agents, software that carries out a chain of steps on its own, to find a way in, take the keys, and pull out data with little human help and at a speed no part-time IT team can match. That changes what a CEO has to decide, and how soon.
What Changed Since Summer
On November 13, the FBI issued an alert about a tool called ARTEX, an automated hacking framework first published as open source in July and used by criminals since at least August. The FBI says it lets people with minimal technical skill run complex attacks against many targets at once, and that some intrusions went from first targeting to privileged access, meaning administrator-level control, in under three hours. One campaign took personal information on about 67,000 people from financial institutions in South Korea. Another went through a technology vendor’s cloud backup service and opened access to more than 2,250 of that vendor’s downstream customer environments.
On November 16, Google’s threat intelligence group reported that attackers are automating the step that used to take skill and time: reading a software maker’s fix, working out what flaw it repairs, and building an attack for every server that has not installed it yet. Google found one such system, set up to run continuously against popular web software, drafting and launching attacks without human intervention.
And on November 17, Reuters reported that Marsh’s clients logged 105 cyber incidents in the last two weeks of October, more than double the 48 they logged in the last two weeks of May. Marsh’s U.S. and Canada cyber practice leader attributed a significant number of them to attackers using AI tools, and described AI not as a new kind of risk but as an amplifier of every risk a company already has.
The cost side has moved too. In September, Gambit Security, a security vendor, described a single operator running AI agents against hundreds of online retailers at a cost of tens of dollars per company, taking more than 600,000 card records. In one case the agent’s own cleanup routine destroyed the victim’s data.
Why the Good News Ran Out
Our earlier piece made one argument we still stand behind: a flaw that is found can be fixed, and a company with a clear rule for what to fix first can treat a flood of disclosures as weather. That filter still works. What has broken is the clock.
That filter still works. What has broken is the clock.
When a vendor ships a fix, the fix itself describes the flaw. That was always true, but turning a fix into a working attack used to take a skilled person days or weeks, and in our experience most attackers did not bother with companies your size. Now a machine does it quickly and cheaply, against everyone who has not updated. The publication of a fix has become the starting gun. Our earlier advice was to give the most dangerous flaws on internet-facing systems “days.” Our view now is that for anything your customers, vendors, or the public can reach from the internet, days is too slow.
Why a Midsize Company Is Now Squarely in Range
In our view, midsize companies used to benefit from being not worth the effort. A skilled attacker’s time was expensive, so it went to large targets with large payouts. AI removes that protection. When an attack costs tens of dollars and runs unattended, the attacker no longer chooses targets; the attacker sweeps everything and keeps what works.
When an attack costs tens of dollars and runs unattended, the attacker no longer chooses targets; the attacker sweeps everything and keeps what works.
In our judgment, three traits of a typical midsize company make the sweep more likely to work. The IT team is small and works business hours, while a three-hour intrusion fits easily inside a weekend. Outside firms, such as the managed service provider, the backup service, and software vendors, hold standing access, and the FBI’s backup-vendor case shows one compromised vendor can open thousands of customers at once. And nobody has a complete list of what the company has exposed to the internet, which is the first thing these tools map.
Three Decisions That Belong to You
None of these is a technical decision. Each is a choice about spending, risk, and who is accountable, which is why it lands on your desk.
The first is how fast your internet-facing systems get fixed. Decide that anything reachable from the internet either updates itself automatically or sits with a provider whose contract commits to applying critical fixes the same day. Anything that cannot meet that standard should come off the internet or sit behind a protected connection. Some systems will need an exception, and the exception should be a named person accepting the risk in writing, as we recommended last month.
The second is who is watching when your team is not. Decide whether your company has someone, inside or outside, who sees an intrusion at two in the morning and has the authority to shut a door. For most midsize companies that means an outside monitoring service that works around the clock. In our experience it costs far less than staffing the same coverage internally, but it is still a real budget line, and it is one to decide on purpose rather than discover after a weekend. Thanksgiving is next week, and we wrote recently about why holidays matter in Some Ransomware Crews Wipe the Backups Too.
The third is which outside firms hold keys to your systems. Ask for a list of every vendor with standing access, what each one can reach, and whether each login requires a second factor. Then decide which access is still needed, and require in your contracts that vendors tell you within hours, not weeks, when they are breached. We made the same point about field networks in energy in Who Owns the Modem on Your Well Pad?, and it applies to every industry.
What to Tell Your Board and Your Insurer
Your board does not need the names of the tools. It needs to hear that the risk changed in speed, not only in volume, and what you are doing about it. Two numbers from our earlier piece still belong in the report: how many must-fix items are open and how old the oldest one is. Add two more. How many hours does it take us to fix a critical flaw on an internet-facing system? And is anyone watching after hours, yes or no?
Your insurer is asking similar questions. GlobalData, a research firm, expects global cyber insurance premiums to rise 6.9 percent in 2026 and 10 percent in 2027 as attacks become more frequent, and underwriters price what they can see. We covered what is negotiable at renewal in U.S. Cyber Insurance Stopped Getting Cheaper. A company that can show fast updates, round-the-clock monitoring, and a clean vendor access list has something to negotiate with.
A company that can show fast updates, round-the-clock monitoring, and a clean vendor access list has something to negotiate with.
Do You Need AI to Fight AI?
You will hear that the only answer to machine-speed attacks is machine-speed defense, and there is truth in it. Google makes that argument, and it also sells security services, so weigh it accordingly. Our view is more modest. A midsize company does not need to buy an AI security product to close most of this gap. It needs automatic updating, a monitoring service that already uses automation, fewer doors open to the internet, and fewer vendors holding keys. Those are decisions and contracts, not research projects. If AI-assisted defense earns a place later, it should go through the same governance as any other AI your company uses, the discipline we describe in our piece on ungoverned AI agents.
What This Will Not Do
The FBI’s cases describe specific campaigns against sectors from finance and government to manufacturing and energy, and the alert does not discuss company size. Marsh’s figures come from its own clients over two short windows, and the share it attributes to AI is the firm’s estimate. Gambit is a security vendor describing one operator. Google’s report describes what attackers are building, not how often they succeed. The insurance forecast is global, not specific to midsize companies. Our view that a midsize company should now measure fixes on internet-facing systems in hours, and our cost comparison for outside monitoring, are our judgment. None of the three decisions will stop a determined, well-funded attacker, and none replaces the basics we described last month. What they do is close the gap that these new tools are built to find.
Where to Start
This week, before the holiday, ask your IT lead three questions. Which of our systems face the internet, and how fast does each one get a critical fix? Who is watching our systems on Thanksgiving weekend, and what are they allowed to do if they see something? Which outside firms hold standing access, and when did we last check it? If the answers are slow or vague, start there. If you would like an experienced CIO to work through these decisions with you, that is the kind of question our CIO IQ advisory is built for, and our cybersecurity work picks up where the decisions need hands. If you would rather talk it through first, you can reach us through the contact form on our About page.


