A Rough Week in Cybersecurity: AI’s Role

October is Cybersecurity Awareness Month, and this week made the case on its own. On Wednesday, TechCrunch said the internet was having “a very bad cybersecurity day,” and it had a list to prove it. Between Saturday and Wednesday a certificate authority’s keys turned out to be guessable, PHP shipped emergency fixes, Node.js patched a remote code execution bug, and a widely used AI tool was hijacked for a few hours. Three of the biggest stories in the cybersecurity news this week involved AI, each in a different way: AI found the flaw, AI ran the attack, or AI software was the target.

I build and look after websites and Shopify stores for small businesses, so I read all of this with one question in mind: does it reach an ordinary company site? Some of it does. Here’s what happened, where AI came into it, and what I’d check before the weekend.

Which of this week’s cyber attacks involved AI?

PHP2Shell was found with AI. The two critical PHP flaws that led 211 Japanese organisations to pause or limit services by October 22 (I covered who they affect in PHP2Shell: Who the Critical PHP Flaw Affects) were reported by Wiz Research, OpenAI’s Codex Security team and the security firm Calif. Wiz says it used its own Wiz Atlas plus Google’s Gemini 4 Argon and OpenAI’s GPT-5.6 Sol to find the bugs, build working exploits and test the fix. The faulty code had been in PHP for more than a decade, back to version 5.4. By Wiz’s account, AI let one researcher finish in a week what used to take a top team months.

AI agents ran a break-in campaign. On October 21, Microsoft said a China-aligned group it calls Storm-0940 had been using ARTEX since at least September. ARTEX is an open-source penetration testing tool that can run on any AI model, and its agents work through an attack from scouting to stealing data with little human help. The group tried to sign in to more than 2,250 Microsoft 365 tenants (a tenant is one organisation’s account), and one password-spraying run hit 2,900 failed sign-ins a minute. Most attempts went nowhere. In several cases, though, Microsoft saw the attackers get from first target to privileged access in under three hours, and in one an app they controlled made itself a global administrator.

The way in matters for small businesses. Microsoft ties Storm-0940 to the ConnectWise Cloud Backup breach that ConnectWise disclosed on October 16. IT providers use that service to back up their clients’ Microsoft 365 accounts, and at least 173 organisations were reached through it, according to Check Point’s weekly report. The attackers came in through an app their victims had already trusted with access.

An AI tool was hijacked. LiteLLM is open-source software that sits between an app and the AI models it calls, often holding the API keys for all of them. On October 21, someone holding stolen developer credentials published a malicious version, 1.107.0, built from a lookalike copy of the project. It was on PyPI, the Python package registry, from 11:39 to 13:23 UTC, and behind LiteLLM’s latest Docker images until 14:20. It was built to steal cloud credentials, SSH keys and AI provider keys, and to install a loader that could fetch more code later. LiteLLM recommends pinning 1.106.2, the last clean release before this one.

A smaller one: OpenCode, an AI coding assistant, said on Saturday, October 17, that a breach had affected 1.4% of its Console users, all of them people who had linked outside AI providers such as Anthropic, OpenAI or Google.

What else was in the cybersecurity news this week?

eMudhra, an Indian certificate authority, had eight intermediate certificates whose private keys could be recovered from public data. Mozilla got the report on October 17. The researcher Scott Helme showed that one key was simply the number 4, and recovered others in under a second on his laptop. With those keys, anyone could have issued a working HTTPS certificate for any website. eMudhra has revoked the certificates. Firefox and Chrome blocked them on October 17 without needing a browser update, Apple followed on October 19, and eMudhra says it found no sign they were used to issue certificates to end users.

Node.js released fixes on October 21 for ten flaws, four of them rated high. The worst is an out-of-bounds write in its Brotli decoder that allows remote code execution, and two more let a connection slip past TLS hostname checks. The fixed versions are 22.23.4, 24.22.1 and 26.12.1, and the same update stops trusting the eMudhra certificates.

What should you check this week?

Most of this week’s fixes land on whoever runs the server. Shopify stores carry less of that, because Shopify runs and patches its own servers (and they don’t run PHP, which is why store owners had nothing to do for PHP2Shell). Their apps, staff accounts and connected tools are still theirs to check. If you run your own site, this is my list for the next few days:

• If your site runs on PHP (WordPress, WooCommerce, EC-CUBE, Laravel), confirm the server runs PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35, or a host build with the fix.
• If an IT provider manages your Microsoft 365, ask whether they used ConnectWise Cloud Backup and whether they’ve reviewed sign-ins since at least October 15, and app sign-ins back to October 1 where the logs allow, which is what Microsoft advises.
• If your site has an AI chatbot or another AI feature, ask your developer whether anything installed LiteLLM 1.107.0 on October 21. If it did, the fix is a clean rebuild and new keys for everything that machine could reach.
• If your site runs on your own Node.js server, update to 22.23.4, 24.22.1 or 26.12.1.

Will AI make website security worse?

For sites that get patched slowly, yes, and soon. You can already see it in the numbers. Microsoft fixed 1,413 vulnerabilities in its October updates, more than in all of 2025, and SecurityWeek puts the surge largely down to AI finding bugs. PHP2Shell sat for over ten years in the language behind about seven in ten websites until AI-assisted researchers found it. Storm-0940 used AI agents to get a break-in down to a few hours.

I expect this to keep going as the models improve. The US government has even renamed AI Super Intelligence in its own documents, and whatever you call it, it’s getting very good at reading old code. Expect more decade-old bugs to surface in software everyone has trusted for years: server languages, content management systems, plugins and the libraries underneath them. Some will be found by researchers who report them first. Some won’t.

The same models help defenders. Wiz used them to check the PHP fix, and every bug a researcher finds and gets patched is one fewer for an attacker. What changes is the clock. PHP2Shell drew attack attempts hours before the fixes were out, after the CVE numbers leaked the day before. A site that gets updated whenever someone remembers is the one most likely to get hit.

What does a maintained and monitored website look like?

Nothing exotic. Someone you can name applies security updates to the server, the CMS, the plugins and the libraries, within a day or two for critical fixes, and then checks that the new version is the one actually running.

That person also hears about it when things change: when the site goes down, when a new admin account appears, when files change that shouldn’t, when the HTTPS certificate is about to expire or comes from an issuer browsers no longer trust. They keep a short list of every app, plugin and integration with access to the site or its accounts, and remove the ones nobody uses. ConnectWise and LiteLLM were both trusted tools that ended up working against the people using them, and when a tool like that is breached, its keys get changed the same day.

If you’d rather someone else kept up those habits, priority security patches and security monitoring are part of my care plans.

Common questions