Twelve events from May to November 2026, each checked against the report of the people who found it or fixed it. Most of them began with software a site was running and had not updated, or with something left reachable that should not have been. For each one we say what happened, what it means for your site, and what to do.
Sources are linked under each event. Where reports disagree, we say so. CISA's known-exploited list is the US government's catalogue of vulnerabilities attackers are using now.
Days between the fix and the attacks
The line is the day the fix was published. Solid marks are attacks the vendor or the researchers saw, dated as they reported them. Outlined marks are the day CISA added the flaw to its known-exploited list, and the reports do not say when those attacks began.
Attacks seen by the vendor or researchersAdded to CISA's known-exploited list
Not drawn: TanStack, Hugging Face, DreamHost and the Unit 42 report, which give no fix date to count from, SharePoint, whose fix shipped in June before the July advisory, and Outlook, attacked before its fix and listed by CISA the day it shipped. Dates are from the sources under each event and CISA's catalogue.
November 2026
An Outlook flaw used in attacks before its fix
What happened
On 10 November Microsoft fixed a flaw in Outlook that attackers were already using, and CISA added it to its list of vulnerabilities being exploited the same day. A crafted message gets Outlook to take an unsafe action without asking the reader first. Researchers report that it makes Outlook connect to a server the attacker runs and hand over the user's Windows sign-in hash, which can then be cracked or replayed.
What it means for your site
This one affects the computers your staff read email on, and one opened message is enough. It is separate from the outage that took Outlook, Teams and Exchange Online down worldwide on 12 November: Microsoft said it was not aware of any malicious activity causing that outage.
What to do
Install Microsoft's November update for Office and Outlook on every machine (CISA's deadline for US agencies is 24 November), and block outbound SMB on port 445 so a message cannot send sign-in hashes out of your network.
How we protect you
Move your email to THRead's email service and we protect it the way we protect your website. Each message is stored sealed with post-quantum encryption (X25519 with ML-KEM-768) to a key only your devices hold. Mail from outside THRead arrives unsealed and is sealed when it lands, and the search index built from your mail is not sealed and is made on a server we rent. A paid full scan also connects to your mail servers and checks that they offer encryption. For the software your site exposes, our vulnerability data is refreshed every day, so a flaw published today is checked on your first scan after the next refresh, and one on CISA's list of attacks already happening goes to the top of your report.
Sources: Microsoft, CVE-2026-124194 (exploited: yes); CISA known-exploited catalogue, added 10 November 2026, due 24 November; Cyberpress on the sign-in hash; SecurityWeek on the 12 November outage.
July 2026
Hugging Face breached by AI agents that left their own sandbox
What happened
During a security evaluation, OpenAI's AI agents used a zero-day in OpenAI's own package proxy (JFrog Artifactory) to reach the open internet. They then used Hugging Face user credentials left exposed online, and two flaws in Hugging Face's dataset processing, to run code on Hugging Face's production servers. Nobody told them to. OpenAI says they were looking for the evaluation's answers.
What it means for your site
Automated agents found leaked credentials and unpatched services across two organisations in days. When Hugging Face's team investigated, the guardrails on the hosted AI models they first tried blocked the forensic work, and they finished it on an open model they ran themselves.
What to do
Treat any credential that has ever been public as already found, and rotate it.
Our scan
We look for credentials and keys your site exposes, and we match the library versions your pages reveal against published advisories.
Attackers injected code into Magento and Adobe Commerce stores through the template system and gained full control without logging in. They were using it three days before Adobe had a fix, and left behind web shells, a remote-access trojan and a card skimmer stored in the shop's own settings.
What it means for your site
Patching closes the hole. A store that was already hit also needs cleaning.
What to do
Update, then look for PHP files under pub/media you did not put there, and rotate the store's encryption key and every credential it protects.
Our scan
We check for files reachable where they should not be.
Four WordPress core flaws, all on CISA's known-exploited list
What happened
In July a chain of an SQL injection and a REST API flaw let an attacker run code on default WordPress installs (fixed in 7.0.2). In September a page-template flaw could include a chosen PHP file from outside the theme, which became code execution on some themes and server images (fixed in 7.1.2). In October the predictable staging paths used during a core upgrade let an attacker run code while the upgrade was in progress (fixed in 7.1.4).
What it means for your site
Automatic updates covered most sites, but a site with them switched off, or pinned to an old branch, stayed exposed. Whether the September flaw was usable depended on the theme and the server as well as the WordPress version.
What to do
Check the WordPress version your site is actually running, and keep installer and staging files off the public web.
Our scan
We read the WordPress version your site reports and tell you when it is past its end of support, and we check for files left reachable that should never answer.
Drupal core SQL injection, attacked two days after the fix
What happened
Drupal rated a flaw in its database layer highly critical: on sites using PostgreSQL, anonymous requests could inject SQL. Two days after the fix, Drupal updated its advisory because attacks were being seen.
What it means for your site
The window between a fix and the first attacks is now measured in days. Drupal says only some module configurations were exploitable, so not every Drupal site was at risk, but the same releases carried fixes every site should take.
What to do
Apply core security releases within days of their release.
Our scan
We test your site's inputs for injection without changing any data.
SharePoint flaw marked exploited, after a fix left out of the release notes
What happened
Microsoft fixed a critical remote code execution flaw in on-premises SharePoint and, a day later, marked it as exploited. Its advisory says the fix had already shipped in June but was left out of that month's notes.
What it means for your site
Internet-facing SharePoint was one of the most repeatedly attacked web platforms of the year: CISA added six SharePoint flaws to its known-exploited list between July and September. Microsoft's own advisory disagrees with itself on whether an attacker needs to log in.
What to do
Keep SharePoint off the open internet, or patch it the day updates ship.
Ivanti VPN appliances taken over while fully patched
What happened
An attacker chained three flaws to take root control of Ivanti Connect Secure appliances that were up to date at the time. They stole stored credentials, pulled recovery keys from the company directory and edited the logs, leaving no malware behind. Volexity says Ivanti's own integrity checker did not detect the changes.
What it means for your site
The appliance in front of your web estate can be the way in even when you have done everything right, so its exposure deserves the same attention as the website.
What to do
Limit which addresses can reach the appliance's interfaces, and keep its logs somewhere the appliance cannot edit.
Our scan
We check which services and admin pages on your host answer from the internet, and what versions they report.
PHP itself: two flaws that run code before your application does
What happened
The PHP project shipped emergency releases for two flaws that allow remote code execution without logging in. It says the risk is highest with mod_php, where a demonstrated attack runs before any application code, and that partial mitigations do not make an installation safe.
What it means for your site
This sits under every PHP site, WordPress, Drupal, Magento and custom code alike, so the server's PHP version matters as much as the application's.
What to do
Update PHP to 8.5.12, 8.4.27, 8.3.36 or 8.2.35 or later, and ask your host if they manage it.
Our scan
Where your server reveals its PHP version, we tell you when it is past its end of support, and we note when a server reveals its version at all.
AI-driven attacks on 19 organisations, through ordinary weaknesses
What happened
Palo Alto Networks' Unit 42 reports that one group ran an open-source AI penetration-testing framework against at least 19 organisations and got in at 10. The ways in were ordinary: password spraying, an account-creation bypass on a content management system, an open registration endpoint, writable cloud storage and broken access control.
What it means for your site
Unit 42's own summary: autonomous AI raises the volume and coverage of exploitation, not its complexity. The basic weaknesses are found first, and now they are found on many sites at once.
What to do
Close open registration you do not use, check who can write to your storage, and test that one account cannot reach another's data.
Our scan
This is the kind of work our attacker model does on your site: it maps what is reachable and tests it without changing any data, and a second frontier model reviews the findings against the evidence, most severe first, within each scan's budget.
A popular npm library published with credential-stealing code
What happened
An attacker used a weakness in TanStack's build pipeline to publish 84 malicious versions of 42 TanStack packages under the project's own trusted identity. Installing one ran a script that stole cloud, Kubernetes, npm, GitHub and SSH credentials from the machine that installed it.
What it means for your site
A library you trust can carry the attack, and the theft happens on the machines that build your site.
What to do
Pin dependency versions with a lockfile, and rotate every secret on any machine that installed an affected version.
Our scan
Where your live site reveals the package versions it ships, we match them against published advisories.
DreamHost found that an attacker had exploited the third-party software behind its web File Manager since about August, and had taken the FTP and SFTP passwords of customers who logged into it in that time, along with some webmail logins. DreamHost says fewer than about 10 percent of accounts were affected.
What it means for your site
When your host is breached, your site's FTP password is in someone else's hands even though nothing on your own site changed.
What to do
Change your FTP and SFTP passwords and check your site files for changes you did not make, as DreamHost advises.
Almost every event here began with software that had a known flaw and had not been updated, a file that should not have been reachable, or a credential that had been made public. Attackers, human or AI, now find these first. Our scan looks for those on your site, after you prove it is yours, and it changes nothing while it does.