Each problem began in software or a service that businesses use but did not build.
Between 17 and 21 October 2026, five serious security problems came to light. Each one began in software or a service that many businesses use but did not build: the software behind most websites, a place where companies keep their code, a tool installed from outside, a platform that stores passwords, and a company that issues website certificates.
On 21 October, the PHP project released fixes for two critical flaws, CVE-2026-109618 and CVE-2026-109619. Both are rated 9.2 out of 10. On a website that has not been updated, an attacker could run their own code on the server without logging in. PHP runs 69.3% of websites whose server-side language is known, including every WordPress site.
Details leaked on 20 October, so the PHP project moved its release forward by a day. Cloudflare saw the first attacks at 3:43 pm Singapore time on 21 October. That was about six hours before the fix was published, according to Sansec. The Canadian and Dutch governments took some of their websites offline as a precaution. The UK NCSC advised taking a system offline if it cannot be fixed.
The fixed versions are PHP 8.2.35, 8.3.36, 8.4.27 and 8.5.12. The PHP project no longer updates PHP 8.1 or older. Security support for PHP 8.2 ends on 31 December 2026.
What an attacker could do with your website. An attacker's code on the server could change the files the website itself can change. On a compromised site, an attacker could:
On 21 October, GitLab fixed CVE-2026-110264, rated 10.0 out of 10, the highest possible score. On a self-managed GitLab server, an attacker could run code without logging in. Every version from 7.0, released in 2014, is affected. With that access, an attacker could copy your code, or change it so they can get back in later. That change could then ship with your next release.
The fixed versions are 19.5.1, 19.4.3, 19.3.6, 19.2.9, 19.1.9, 19.0.10 and 18.11.13. If you updated for the earlier flaw fixed on 12 October, update again. Customers of GitLab.com do not need to act.
On 21 October, a malicious version of LiteLLM, 1.107.0, was published on PyPI and on Docker Hub, where it also carried the "latest" tag. It was built from a lookalike of the project's code, LiteLLM says. Once installed, it ran every time Python started, even if no one used the tool. It collected passwords and keys, including SSH keys, cloud logins and AI provider keys. Stolen AI and cloud keys can be used to run services on the victim's account, leaving the victim with the bill.
This is the second compromise of LiteLLM this year, after versions 1.82.7 and 1.82.8 in March. Treat any machine that installed 1.107.0 as compromised. Rebuild it, and change every password and key that machine could access. LiteLLM names 1.106.2 as the last clean version.
On 21 October, Railway disclosed that someone had reached the systems that store its customers' settings, including their passwords and keys. Some of that data may have been read in decrypted form. These settings can include the logins for a company's databases. Anyone holding them could then open the data behind its apps. Railway paused all deployments and told every customer to change all stored passwords, API keys and registry logins.
Railway believes the attack was part of a series aimed at platform companies, with the goal of reaching AI providers.
Certificate authorities issue the certificates behind the padlock in your browser. They sign them with secret keys that should be random numbers over a hundred digits long. On 17 October, a researcher found that eight of eMudhra's keys were numbers below 256. One of them was the number 4. Anyone could have worked out the keys and made fake website certificates that many devices would have trusted.
Browsers blocked the eight keys within days: Firefox and Chrome on 17 October, Apple on 19 October and Microsoft on 20 October, per Mozilla. Java keeps its own list of trusted keys, and was only updated on 21 October (Oracle, OpenJDK). eMudhra says it issued no certificates from the eight keys, and there is no public evidence that anyone misused them.
A researcher at Calif found the PHP flaws with help from an AI model. OpenAI and Wiz reported them independently. In 2025, the systems in DARPA's AI Cyber Challenge found 18 real flaws in open-source software. Expect more urgent updates, including for old code.
The UK NCSC wrote in 2025 that the time between disclosure and exploitation has shrunk to days, and that AI will almost certainly shorten it further. This week, the first PHP attacks came before the fix.
LiteLLM uses keys to connect apps to AI providers. Railway believes its attackers were aiming at AI providers. Check which passwords and keys each AI service can use. Treat it like any other supplier.
Ask your web developer or hosting company: have you installed the PHP fixes released on 21 October? Ask for the answer in writing. Some hosting companies apply the fixes to their own builds, so a different version number can still be safe. Ask them to confirm it.
This covers web browsers, computers and phones, Java and anything that includes it, and any GitLab server you run yourself. If your developers use containers, rebuild them to include the updated list of trusted certificates.
If you use Railway, change every password and key stored there. If anyone in your team installed LiteLLM 1.107.0, rebuild that machine first, then change every password and key that machine could access.
Record which systems and data each one can access. Name the person who keeps it updated. Use the list to check whether a security advisory affects you.
These are the basics that the CSA's Cyber Essentials mark covers: knowing your hardware and software, keeping it updated, and controlling access. If your business holds a Cyber Essentials or Cyber Trust mark, file your provider's written confirmation with your software update records.
A leaked brochure claims a police forensics tool can unlock some iPhones from about 8 metres away. The claim is unverified. Apple fixed several Bluetooth flaws in iOS 27, so keep phones on the latest version.
Each claim above links to its source, read on 21 and 22 October 2026.