Website security monitoring now has to work in hours, not at the next audit. In November 2026, Google's Threat Intelligence Group reported finding an attacker's AI framework that reads each new software patch, works out the flaw it fixes, writes an exploit and runs it, with no person involved. Once a fix is published, the fix itself is the map for the attack.
That is why every website in Studio, our service for running a website from Claude or Codex, is now checked for security every hour, and again within five minutes of any change. This post covers what Google found, the attacks websites are facing, and how the watch works.
What did Google find?
Google's AI Threat Tracker for November 2026 (opens in a new tab) describes a framework its researchers named exploit_agent, found on an exposed staging server. It connects to several AI models and keeps comparing popular web application software before and after each patch. When it spots the likely flaw, it drafts an exploit and runs it automatically. Google calls it the first case it has seen of AI fully automating vulnerability discovery and exploitation "without a human in the loop".
The targets were websites. The operator appeared most interested in French sites in media and entertainment, and in education.
The same report gives the wider picture:
- Exploits for flaws that had no fix yet (zero-days) rose from an average of 8 a month in 2025 to 11 a month in 2026, then to 22 in August.
- Disclosures and exploits together nearly doubled between January and August 2026.
- In one case, researchers using AI had working exploits for two newly disclosed PaperCut flaws within a day.
- Attackers are tampering with the settings of AI coding assistants so that a backdoor is dropped when the assistant runs on a developer's machine.
For anyone who owns a website, the point is simple. The gap between "a fix is out" and "attackers are using it" is shrinking towards a day. A site that is updated once a quarter, or when someone remembers, is exposed for most of that time.
Which attacks are websites facing?
These are the five kinds of attack that come up most in the public record, and what each one means for a website.
1. Known flaws in the software a site is built from
A modern website is built from hundreds of open-source packages. When a flaw in one of them is published, every site on that version is exposed until it updates.
React2Shell is the clearest recent example. It was disclosed on 3 December 2025, rated 10 out of 10 for severity, and let anyone run their own code on servers using React Server Components, including Next.js sites built with the App Router. Attacks were seen within two days, and the US Cybersecurity and Infrastructure Security Agency (CISA) added it to its list of flaws known to be used in attacks on 5 December (Rapid7 (opens in a new tab), Wiz (opens in a new tab)).
2. Poisoned packages: software supply chain attacks
Sometimes the package itself is the attack. Someone takes over a maintainer's account and publishes a version with malicious code inside, and it reaches every project that installs it.
- On 19 May 2026, malicious versions of 323 npm packages were published from a single compromised account. The code stole credentials, and researchers reported that it also went after AI coding tools (Expert Insights (opens in a new tab)).
- In August 2026, the ChainDrop worm started from one maintainer's hacked GitHub account. Aikido counted at least 868 affected packages across 1,381 versions. It stole cloud and build-system secrets, and because it published through the projects' own release workflows, the bad versions came with valid provenance records and looked legitimate (BleepingComputer (opens in a new tab)).
These attacks mostly hit the computers and build systems that install packages, not the live site directly. The defence is knowing exactly which version of every package a site uses, updating deliberately rather than blindly, and never running a package's own install scripts while preparing an update.
3. Files and keys left in the open
Some of the most damaging attacks need no flaw at all. In 2024, Palo Alto Networks' Unit 42 described an extortion campaign that scanned 110,000 domains for exposed .env files, the files that hold a website's passwords and keys. It collected more than 90,000 unique values, about 7,000 of them for cloud services, and used them to break in and demand payment (Unit 42 (opens in a new tab)).
Keys also leak through code. GitGuardian's 2026 report counted about 29 million new secrets exposed on public GitHub in 2025, 34% more than the year before (GitGuardian (opens in a new tab)).
4. Attacks aimed at AI coding assistants
As more website work happens through AI assistants, the assistants become a target. Google's report describes backdoors planted through coding assistant settings, and the May npm attack went after AI coding tools on developers' machines. An AI assistant that could publish to your website by itself would be well worth taking over.
5. The basics that still get missed
- Security headers. These are instructions a site sends to browsers: always use https, only run approved scripts, don't guess file types, and don't let other sites frame the page, a trick used to hijack clicks.
- Saying too much. A site that announces which framework it runs tells attackers which exploits to try.
- Certificates. Since 15 March 2026, public TLS certificates last at most 200 days, falling to 100 days in March 2027 and 47 days in March 2029 (SSL.com (opens in a new tab)). More renewals mean more chances for one to lapse and put a browser warning in front of every visitor.
How Studio's security watch works
Studio connects a website to Claude or Codex so a team can write, approve and publish changes in plain words. Security is part of it, and it runs whether anyone asks or not.
When it checks. Every site is checked every hour, within five minutes of any change going live, and whenever our engineers ask. A new advisory for something a site uses is caught by the next hourly check.
What it checks.
| Check | What Studio looks at |
|---|---|
| What the site is built from | Every package and exact version in the site's lockfiles, looked up in the open vulnerability database (which includes GitHub's advisories) and in CISA's list of flaws known to be used in attacks |
| How the site looks from outside | The certificate's end date, the security headers, whether the site announces what it runs on, and, once a day, whether /.env or /.git/config can be opened |
| Keys in the code | A committed .env file with real values, or a private key file. The values are never copied into a report |
Who is told. When something is vital, such as a critical flaw in a package that reaches the live site, a flaw on CISA's list, an exposed file or a certificate about to lapse, our engineers are alerted at once. Everything else is listed for them and summarised each morning.
How it is fixed. Where it can, Studio prepares the update on a separate branch of the site's code. It changes only the package list, and it installs nothing and runs no package's code while doing so. One of our engineers approves it, and only then is it published. Anything Studio can't prepare, such as a move to a new major version, goes to an engineer. Nothing is merged automatically, and every approval is recorded.
What you see. You are not sent alarms. Your website's page in Studio shows that it is watched and when it was last checked. Once something is fixed, it shows a one-line note of what was done, such as "Security update applied: next". Your AI assistant can tell you the same.
What leaves Studio. Only package names and versions, sent to public databases to look them up. Nothing about you, your site or your code.
Is it safe to let AI change your website?
After point 4, it's a fair question. Studio is built so that an AI assistant can draft but never publish on its own:
- Every change is previewed, and nothing goes live until a person with the right role approves it.
- Each person has a role: who can draft, who can approve and who can publish.
- Studio tells your AI assistant to treat words found on pages, in search terms and in comments as content, never as instructions.
- Every action is recorded, and a published change can be undone.
- The code stays in your own GitHub account, and Studio can only read and write the contents of the repositories you give it.
For the wider controls an AI system needs, see our AI security checklist for LLM apps and our guide to MCP and agent tooling.
What website security monitoring doesn't replace
Monitoring finds known problems quickly. It is not a penetration test. It doesn't find flaws in a site's own code, problems in a hosting account or domain settings, or flaws nobody has published yet. A match between a package version and an advisory also isn't proof the site can be attacked: an engineer decides whether the affected part is in use.
If your site was built quickly with AI tools, our post on the vibe coding gap covers the checks to run before launch.
A short checklist for any website owner
- Know what your site is built from, down to the version of each package.
- Find out within hours, not at the next audit, when one of them has a published flaw.
- Keep keys out of your code, and make sure
.envand.gitcan't be opened from the web. - Send security headers, and don't announce your framework.
- Watch your certificate's end date, which now comes round at least twice a year.
- Make every change, including changes an AI assistant drafts, go through a person's approval.
To see where your site stands today, start with our free website audit, which includes a security check. To have your site watched and kept up to date, see Studio.
Frequently asked questions
What is website security monitoring?
Website security monitoring means checking a site on a schedule for known security problems: flaws in the software it is built from, exposed files and keys, missing security headers and certificates about to lapse. Good monitoring checks often, alerts someone who can act, and ends in a fix.
How often should a website be checked for security problems?
At least daily, and ideally every hour. Google reported in November 2026 that attackers are using AI to turn published patches into working exploits automatically, so the time between a fix and an attack is now counted in days or less.
What is a software supply chain attack?
It is an attack carried out through the software a site depends on, rather than the site itself. An attacker takes over a package maintainer's account and publishes a version with malicious code inside, which reaches everyone who installs it.
Will Studio send us security alerts?
No. Alerts go to Agitech's engineers, who fix the problem. You see that your site is watched, when it was last checked, and a short note of each fix once it is done.
Does Studio work with sites Agitech didn't build?
Yes. Studio works with any website whose code is in GitHub. Today the security watch reads JavaScript and PHP lockfiles, and it checks every site from the outside, whatever it is built with.