Did AI Just Break the Internet?

Between January 1 and October 23, 2026, 86,292 software security flaws were made public. That is 2.2 times as many as in the same months of 2025, and more than all of 2025 put together. Over the same stretch, the U.S. government’s list of flaws attackers have been caught using grew by about a third.

2.2xas many security flaws made public January 1 to October 23, 2026, as in the same months of 2025CyberHire count of the CVE List, 2026

1.3xas many flaws confirmed in use by attackers, January 1 to October 27, 2026CISA catalog, RLK count, 2026

Then came the week of October 19. On October 21, PHP, the code behind most websites, shipped fixes for two flaws that had sat in it for more than a decade. Wiz Research found them in August with help from AI, and wrote that one researcher did in a week what used to take a world-class team months. By that afternoon the Metropolitan Police, the Edinburgh and Oxford councils, the Dutch government’s DigiD login service and the Port of Rotterdam had taken websites offline to patch. The same week brought a poisoned AI tool, a company that vouches for websites with master keys close to all zeros, and a top-severity flaw in GitLab. The internet called it Black October.

My read: AI did not break the internet. It flooded the inbox. Twice as many flaws were announced, only a third more were used, and the risk is a security team that reads every alert in order while the few dangerous ones sit in the queue. If your board has asked whether you are exposed, the Friday check and a draft reply are near the end.

What happened, in plain English

Each publicly known software flaw gets an ID number called a CVE. CISA, the U.S. government’s cybersecurity agency, keeps a list of the flaws it has confirmed attackers are using. I downloaded that list and counted it, and most of the comparisons below come from it.

The table runs from most likely to touch a typical company to least, as of October 28, 2026. Open “More” on any row for how widespread it is, what it did, and how big a deal it is. The last two rows made headlines, and most companies can skip them.

What it is	What happened	Does it touch you?
PHP2Shelltwo flaws in PHP, the code most websites run on	Anyone on the internet could take over a web server without a password. Researchers found the flaws using AI.	Probably. If you have a website, online store or customer portal, ask IT to confirm it is updated and restarted. Versions 8.1 and older got no fix and need replacing.
More: how widespread, what it did, how big a deal How widespreadNearly every version of PHP from the last decade. About 5.1 million servers exposed online, by one estimate. What it didAttackers were probing before the fix existed. Police, council and Dutch government sites went offline to patch. How big a deal is this?The biggest item of the month. The closest comparison is Log4Shell in December 2021, a flaw in a building block under thousands of products. Are you an idiot for running PHP? No. About 69 percent of websites use it. The mistake is running a version too old to get fixes.
SafePaya ransomware gang	U.S. and European agencies warned that this gang breaks in with stolen passwords and old flaws in remote-login systems, wipes backups, then locks files for ransom.	Yes, if staff log in from outside the office. Ask three things: does remote login need a second step, like a code on a phone? Are the five old flaws patched? Is one backup kept offline?
More: how widespread, what it did, how big a deal How widespreadAt least 600 victims since August 2024. What it didWarning issued October 20. The five flaws it uses date from 2020 to 2024. How big a deal is this?Old news, most victims. Stopping it takes nothing new.
Palo Altoa flaw in a popular firewall	A flaw in the firewall’s remote-login feature. The fix came out in May, and ransomware gangs are still using it against companies that never installed it.	Only if you use Palo Alto firewalls. Ask IT to confirm the May fix is on every one.
More: how widespread, what it did, how big a deal How widespreadFirewalls with one specific login setting switched on. What it didTwo ransomware gangs have used it since June. How big a deal is this?Routine, and that is the problem. These firewalls have appeared 13 times on the government’s list of flaws attackers use.
Ivantia flaw in a remote-access box	Attackers got into a widely used remote-login device without a password, even on fully updated devices.	Only if you use Ivanti Connect Secure. Update it, then have someone check for signs of a break-in, because these attackers covered their tracks.
More: how widespread, what it did, how big a deal How widespreadOne attacker group, since early October. What it didNo malware left behind, and logs were tampered with. Part of the fix was still coming. How big a deal is this?Not new for this product. It had been on the government’s list of exploited flaws 14 times before. If you still run it, budget for a replacement.
UpdraftPlusstolen keys for a WordPress backup tool	Hackers stole the remote-control keys for a popular website backup plugin and used them to plant hidden back doors on customers’ websites.	Maybe, if your website runs on WordPress. Ask whether it uses UpdraftCentral. If yes, update it and look for admin accounts nobody created.
More: how widespread, what it did, how big a deal How widespreadThe plugin is on more than four million sites. Only sites linked to the vendor’s management service were exposed. What it didAt least 99 different back doors found. How big a deal is this?Easy to miss. There was no software flaw to scan for, so security scanners had nothing to flag.
LiteLLMa poisoned AI tool	Attackers used stolen developer logins to publish a fake update of a popular tool for connecting to AI services. It stole passwords and keys from anyone who installed it.	Only if your teams build with AI. Ask whether anyone installed version 1.107.0 on October 21. If yes, change every password and key that computer could reach.
More: how widespread, what it did, how big a deal How widespreadAvailable for a few hours on October 21. What it didAnyone who installed it is told to assume their keys were stolen. How big a deal is this?A repeat offender. This is the second fake release of the tool in 2026, and the government listed three other flaws in it between May and September. Lock the version and wait a day before updating.
GitLaba top-severity flaw	GitLab, where many software teams keep their code, had a flaw that could let an outsider take over the server.	Only if you run GitLab on your own servers. If you do, update this week. The hosted version is fixed.
More: how widespread, what it did, how big a deal How widespreadEvery self-hosted version since 2014. What it didNo attacks reported as of October 22. How big a deal is this?Serious but routine. GitLab has had five flaws on the government’s list of exploited flaws.
Node2Shella flaw in a file-unzipping tool	The code that unzips compressed web traffic, used inside Node.js and Firefox, could be tricked into running an attacker’s code.	A little. It matters only for apps that unzip data from strangers. Ask developers to update Node.js. Browsers update themselves.
More: how widespread, what it did, how big a deal How widespreadNode.js, Firefox, Thunderbird and Red Hat Linux. What it didConfirmed in attacks October 27, though hard to pull off. How big a deal is this?Familiar. A similar flaw in how browsers open images landed on the same government list in 2023.
Zeabura small hosting company	A service for hosting quick apps was broken into using passwords it stored in plain text. Customers’ keys were copied, and the company is shutting down.	Unlikely. Ask engineers where side projects live and whether any hold real company keys.
More: how widespread, what it did, how big a deal How widespread321,810 accounts. What it didShutting down by December 31, 2026. How big a deal is this?Small company, familiar ending. Its second breach in two months, and its master logins never expired.
eMudhrabroken website certificates	A company that vouches for websites’ identity made eight of its master keys so weak that anyone could have faked a trusted website.	Unlikely. Ask whether any of your website certificates come from eMudhra or emSign.
More: how widespread, what it did, how big a deal How widespreadAbout 30,000 customer certificates need replacing. What it didNo fakes found. Chrome blocked the keys within two hours, and other browsers within days. How big a deal is this?Rare, and handled fast. Browsers dropped a much bigger certificate company, Entrust, in 2024.
Tor relayshijacked privacy-network servers	Attackers took over hundreds of servers in the Tor privacy network and stole passwords from people who logged in through them.	Almost certainly not. It only hit people logging into servers through Tor with a password.
More: how widespread, what it did, how big a deal How widespread333 servers, October 1 to 5. What it didNo victim count published. How big a deal is this?Calm down. It started with one reused password.

Look at how each one started, and a pattern shows up.

Is Black October really Black October?

On volume, yes. The 86,292 flaws made public through October 23 beat the 48,151 for all of 2025, and September 2026 alone produced 14,934. Jerry Gamblin’s independent monthly count lands in the same place, at 72,812 through September 30.

On harm, much less. I counted CISA’s list by the date each flaw was added, using the version dated October 27, 2026. It grew by 276 between January 1 and October 27, against 210 in the same window of 2025. That is up about 31 percent, the most of the four years in the chart, and nowhere near double.

Flaws CISA confirmed attackers are using, January 1 to October 27

Attacks are rising. They are not doubling.

CISA Known Exploited Vulnerabilities catalog, version 2026.10.27, entries by date added, January 1 to October 27 of each year. Count by RLK Consulting.

October itself looked ordinary. CISA added 30 flaws from October 1 to 27, 2026, against 27 in the same days of 2025. The busiest month since at least January 2023 was September 2026, with 43, and nobody gave September a color.

Flaws CISA confirmed attackers are using, added per month, 2026

September was the busy month. October looked ordinary through the 27th.

CISA Known Exploited Vulnerabilities catalog, version 2026.10.27, entries by date added. *October covers October 1 to 27. Count by RLK Consulting.

Two cautions about that chart. CISA adds a flaw only when it has solid evidence of attacks and a clear fix, so the real number is higher. The list also runs behind: the PHP flaws went on it five days after the fixes, and on October 8 CISA added flaws from 2015 and 2016. October’s flaws will keep showing up through December, so the fair verdict on Black October comes after December 31.

Most of what landed that week was old news finally made public. Heath Callahan, writing for Forkast, called it “a disclosure wave, not an attack wave.” Wiz found the PHP flaws on August 19. The weak certificate keys were made on April 30, 2025. The Palo Alto fix came out in May.

What the label gets right is the speed. Rapid7 found that serious flaws were used in attacks three days after going public, on average, and 71 percent of the attacks it tracked in the third quarter came within the first week. PHP was worse. Details leaked early, and Amazon saw attackers going after the flaws four days before the fixes shipped.

3 daysaverage time from a serious flaw going public to attackers using it, third quarter of 2026Rapid7, Q3 2026 threat update

So should everyone calm down? About the volume, mostly. Patching once a month is still fine for the big pile of flaws nobody uses. About the speed, no. Once a month is far too slow for the handful that get used.

The takeaway: twice as many flaws were announced, and a third more were used. Judge your team on how fast it fixes the second group.

Why there are twice as many flaws

The UK’s National Cyber Security Centre saw it coming. On May 1, 2026, its chief technology officer, Ollie Whitehouse, told organizations to get ready for a “vulnerability patch wave”. His reasoning was that AI in skilled hands can dig up old weaknesses at scale, and the agency expected a “forced correction” across every kind of software. AI tools got credit for 769 flaws in 2026, against 34 in the same months of 2025, by CyberHire’s count. October fits: Wiz credits Google’s Gemini AI for the PHP flaws, and Node.js credits OpenAI’s security team for the file-unzipping flaw.

Fixing still runs at human speed. Daniel Stenberg, who maintains curl, a widely used tool for moving data across the internet, wrote in April that security reports were arriving at a little over double the 2025 rate, mostly AI-assisted and mostly good, and that the risk of overload was “still real.” PostgreSQL, a widely used database, pushed its next release to November 19, 2026 because of the volume of fixes. PHP took from August 24, when Wiz reported the flaws, to October 21 to ship.

Finding a decade-old flaw in PHP took one researcher about a week. Shipping the fix took two months.

A lot of the free code under a company’s website and systems is kept up by small teams and volunteers, and their inboxes fill at machine speed. Plan for 2027 to bring more patches, arriving in bursts.

Most of the damage came through stolen passwords

Of the rows in the table with known victims, most started with a password or key attackers had stolen or guessed.

The LiteLLM attackers held three developers’ logins, one belonging to an account owner, and used it to switch off the safeguards that should have blocked the fake release. Zeabur’s intruder walked in with passwords stored in plain text, and the company admitted its master logins “had no expiry, had not been rotated since they were created in January 2024.” The WordPress back doors went out through stolen keys, and Patchstack wrote that “this is not a vulnerability in UpdraftPlus.” The Tor servers fell after one operator reused a password. SafePay, with at least 600 victims, gets in with stolen and guessed passwords and flaws from 2020 to 2024.

Old flaws keep showing up on CISA’s list too.

49of the 276 flaws CISA added to its list in 2026 were made public in 2024 or earlierCISA catalog, RLK count, 2026

Two rows break the pattern. PHP2Shell needed no stolen password, and Volexity caught attackers taking full control of fully updated Ivanti devices without one. Both hit systems anyone on the internet can reach. Ivanti’s was one of the boxes at the edge of a network, like firewalls and remote-login devices, and Ivanti and Palo Alto appear on CISA’s list 52 times between them. The NCSC’s advice from May fits: fix the edge of your network first, then work inward.

The takeaway: passwords that expire, a second login step, and patches applied within weeks would have blunted most of these. All of that is in your control, and none of it waits on a vendor.

The case for panic

The strongest case runs through PHP. The flaws sit in the language itself, so Wiz wrote that “even a simple ‘Hello World’ application” could be at risk. No password was needed, attackers had a head start, and PHP runs in places nobody tracks: a vendor portal, an old marketing site, the admin page of a device.

Two things keep it short of a catastrophe. Every supported version of PHP got its fix the day the flaws went public, with plain guidance from the NCSC to update right away and consider going offline if you couldn’t. And the attack is harder than the headline. Wiz describes chaining several flaws and tuning the attack for each target, which slows down mass attacks without making an unpatched server safe. The servers I would worry about run PHP 8.1 or older, which got no fix at all.

The napkin math on your patch queue

Take a made-up company: a 900-person distributor with a four-person IT infrastructure team. In 2025 the team read about 50 security alerts a month that touched its software, at roughly 30 minutes each to read, check whether it applied, and decide. That is 25 hours a month.

Multiply by 2.2 and the same team faces about 110 alerts a month. At 30 minutes each, that is 55 hours of reading before anyone fixes anything.

The alternative is a five-minute sort on every alert, using three questions. Is it on CISA’s list, or is attack code public? Does it touch something anyone on the internet can reach? Does it work without a password? Say six a month pass all three, and each gets two hours of full attention, testing included. Everything else waits for the normal monthly cycle.

Made-up team, 2026 pace	Read everything	Sort first
Alerts per month	110	110
Minutes each to sort	30	5
Sorting hours a month	55	9
Urgent fixes, two hours each	6, so 12 hours	6, so 12 hours
Total hours a month	67	21

The 46 hours saved are the smaller win. With 55 hours of reading spread across four people who have other jobs, an urgent alert can sit for a week before anyone opens it. Against a three-day average, the sort decides whether you beat the attackers.

The catch is the second question. The team can only answer “does this touch something on the internet?” in five minutes if it keeps a list of what runs where, including the PHP and Node.js tucked inside vendors’ products. Without that list, five minutes turns back into thirty.

What a CISO, or whoever owns security, should do

Start with the exposure check. Each question takes minutes, and each yes comes with its next step.

The Friday Check

Does anything we run on the internet use PHP: the website, WordPress, Magento, a vendor or dealer portal? If yes, confirm it is updated and restarted, and list anything on PHP 8.1 or older for replacement.
Do we run Palo Alto firewalls? If yes, confirm the May fix is on every one.
Do we run Ivanti Connect Secure? If yes, update it and have someone look for signs of a break-in.
Does remote login need a second step, like a code on a phone? Are the five flaws named in the SafePay warning patched? Is one backup kept offline? Any no is the place to start.
Does our WordPress site use UpdraftCentral? If yes, update it and look for admin accounts nobody created.
Did anyone install LiteLLM version 1.107.0? If yes, change every password and key that computer could reach.
Do we run GitLab on our own servers? If yes, update it this week.

Then change the standing rules, because the next month like this is coming.

Fix the dangerous few within 48 hours: flaws on CISA’s list or with public attack code, on systems anyone on the internet can reach, that work without a password. Attackers average about three days, so the target has to beat that. Track the typical time, the median, because a few stragglers distort an average.
Sort by “is it being used, and can they reach it” before the severity score. Google Chrome rated four of its flaws critical in all of 2025 and 210 in 2026 through October 23, per CyberHire. Whether the software or the scoring changed, “critical” alone no longer sorts anything.
Ask every vendor whose product runs PHP or Node.js for a written answer on whether it is patched, by a fixed date.
Put an expiry date on every password and key your systems use, starting with the keys to AI services and admin accounts, and change anything older than a year.
If you build software, require two people to approve a release, so one stolen login cannot ship a poisoned version.
After updating a firewall or remote-login box, check whether someone got in first. The Ivanti attackers covered their tracks, so look beyond the logs.

What to tell the board

If the board has asked about Black October, answer the question it asked, then move it to numbers your team controls. A reply like this fits in one email:

A Reply to the Board

We checked our exposure to the October security alerts by [date]. We were exposed on [number] items, and all are fixed as of [date]. [One sentence on anything still open and its date.] Each quarter from now on, we will report how many days it typically takes us to fix a flaw attackers are using on a system the internet can reach, next to the roughly three days attackers need on average. We will also report how many passwords and keys our systems use that never expire, and how many internet-facing systems run software that no longer gets fixes.

Those numbers fall when the team does the work, and they will say more about the next October than any list of flaws.

My prediction: the NCSC’s patch wave is in its first year. 2027 will bring more announced flaws than 2026, and the share attackers use will keep shrinking. The companies that get hurt will be the ones still reading alerts in the order they arrive.

How RLK Can Help

I help leadership teams turn a week like this into a short list of decisions with owners and dates, and into board reporting built on numbers the team controls. Start a conversation.