// this week in AI

This week in AI: critical PHP flaws, Claude Duet, and AI agents that shop and make phone calls

By Trinity · MariVisionOctober 23, 20267 min read
Ale from MariVision in a large dark newsroom, beneath a wall of holographic screens showing a cracked padlock, a small robot with the morning summary, a shopping cart, and a ringing phone
2026 · 10 · 23 — THIS WEEK IN AIWebsites to patch now, and AI agents that work, shop, and make calls

This week the most urgent story isn't a new model: it's an update to run on the server behind your website. Two critical flaws in PHP, the language that powers nearly every WordPress site, were discovered with help from artificial intelligence.

Then there are agents, the AIs that don't just answer but take action: Anthropic unveils an assistant that works on its own and asks for permission before anything sensitive, Walmart opens its checkout to AI assistants, and Yelp has them call restaurants. In between, a small practical update to ChatGPT. For each one: what happened and why it matters to you.

1 · Two critical PHP flaws: WordPress sites need updating now

Large dark server room with a holographic padlock cracked in red floating near the ceiling, next to the glowing outline of an elephant, a green update arrow, and a small shield

On October 21, the PHP team shipped emergency releases 8.5.12, 8.4.27, 8.3.36, and 8.2.35, fixing two vulnerabilities rated critical (a score of 9.2 out of 10). According to the official announcement, they can let someone run code on the server without any password, and they affect every PHP installation that serves web pages. The PHP team is urging everyone to update immediately.

The headline detail is who found them: according to TechCrunch, researchers at Calif and Wiz got there using AI models (OpenAI's GPT-5.6 Sol and Google's Gemini 4 Argon) and reported them before going public, so the fixes were ready. WordPress clarified that the problem isn't in WordPress but on the server: the fix is updating PHP, together with your host.

Why it matters to you: PHP powers roughly 70% of websites, so almost certainly your showcase website or your WooCommerce store too. Ask your host today which PHP version your site runs and whether it's already been updated. If you're stuck on PHP 8.1 or older, you're the most exposed, because those versions no longer get official updates. If AI helps defenders find flaws, it helps attackers too: a site left untouched for years becomes the easiest target. That's exactly the work covered by website maintenance.

2 · Anthropic unveils Claude Duet, the assistant that works on its own and asks for your OK

Dark office at dawn with two overlapping holographic speech bubbles floating near the ceiling, a glowing envelope, an invoice with a red overdue stamp, a calendar, and a big green approval button

Anthropic introduced Claude Duet, an agent that monitors your connected apps (email, team chat, accounting, calendar), sends you a short morning summary of what matters, rechecks your accounts every few hours, and only reaches out when something needs doing. It can run tasks in the background on its own virtual machine, so your computer and phone can stay off.

The examples Anthropic gives are very concrete: finding unpaid invoices, drafting replies to customers, catching an appointment booked on the wrong day. Before it sends a message, pays, signs, or deletes anything, it shows you what it's about to do and waits for your approval; some actions, like moving money between accounts, are blocked no matter what. For now it's a waitlist preview for Pro and Max subscribers in the United States.

Why it matters to you: this is the kind of assistant that people who work alone or with a small team actually need: one that watches your inbox and deadlines for you and only calls you in when it counts. It isn't available in Italy yet, but the question to ask is already here: which apps would you be willing to connect, and which actions do you always want to approve yourself? It's the same principle behind AI agents for businesses: the machine prepares, the person decides.

3 · ChatGPT adds chats locked with a PIN or Face ID

Dark library with a large holographic chat window floating near the ceiling, sealed by a padlock with a numeric keypad, flanked by a fingerprint and the outline of a smartphone

Since October 20, OpenAI has been rolling out locked chats on every ChatGPT plan, including the free one. A locked conversation only opens with a four-digit PIN or with Face ID and Touch ID, moves to a separate section, and no longer shows up in your history or in search. If memory is turned on, locked chats get a memory of their own.

OpenAI is upfront about the limits: it isn't end-to-end encryption and it isn't a separate vault, locking doesn't change how long your data is kept, and if you reset the PIN, all your locked chats are deleted for good.

Why it matters to you: if you use ChatGPT on the shop phone or computer, which may pass from hand to hand, you can hide conversations that involve your numbers or customer names. But it's a lock on the door, not a safe: what you paste into a chat is still a decision to make beforehand. We covered this in the guide on what you can safely upload to AI.

4 · Walmart opens its checkout to AI agents, Amazon shuts its own

Large dark warehouse with shelves of boxes and a holographic portal floating high above, crossed by small robot shopping carts, next to a price tag and a spending-cap gauge, with a red barrier in the background

On October 21, Walmart launched Agent Gateway: a dedicated entry point that lets approved AI assistants search its online catalog and buy on customers' behalf. It uses MCP, the protocol Anthropic created to connect AIs to outside services. Starting in November, every agent will get its own profile in the customer's account, with two modes: approve every purchase or set a spending cap. Return policies still apply when the agent gets it wrong.

Amazon has taken the opposite path: it has started blocking AI agents that shop on users' behalf, and according to TechCrunch several travel and booking sites are keeping them out too.

Why it matters to you: the big online stores are splitting between those opening their doors to AI assistants and those closing them, and small e-commerce shops will face the same choice. Meanwhile, there's work that pays off either way: product pages with prices, variants, and availability written clearly and kept up to date, because an agent reads what it finds and doesn't ask questions. We cover this on our e-commerce page.

5 · Yelp launches a voice agent that calls the restaurant for you

Empty, dark restaurant with an old holographic telephone floating near the ceiling, a small robot talking into the receiver, a glowing clock showing the wait time, and a text-message bubble

In the Yelp app, restaurant pages now show Check for Me: you enter your party size and your number, and an AI voice agent calls the restaurant to ask how long the wait is, then texts you the answer. It's the customer-side version of the AI voice services Yelp already offers restaurants and service businesses.

Alongside it comes Smart Lead Reply, at $30 a month, which responds to prospective customers and collects their requests according to rules set by the owner. TechCrunch points out the oddest risk: an agent calling another agent, and two machines that can't understand each other.

Why it matters to you: Yelp isn't widely used in Italy, but the direction is clear: the phone at your shop or office will start ringing on behalf of machines too. It pays to have hours, wait times, and availability written down and clear, ready to give to anyone who calls, and to ask yourself whether it makes sense to have an AI answer on your end too, at least after hours.

In short

There's one thing to do right away: ask your host whether PHP has been updated to the October 21 releases, because the two critical flaws affect nearly every WordPress site and anyone stuck on PHP 8.1 or earlier is the most exposed. Claude Duet shows what the assistant of someone who runs their own business will look like: it watches email and deadlines and asks for your OK before acting. Locked ChatGPT chats hide your conversations but don't protect them any better on the servers. And AI agents are starting to shop in online stores and call local businesses: clear catalogs and ready-to-go information become an advantage.

Sources: PHP.net (critical security release, October 21) · TechCrunch (PHP flaws found with AI) · Anthropic (Introducing Claude Duet) · TechCrunch (Claude Duet) · TechCrunch (locked chats in ChatGPT) · TechCrunch (Walmart and Amazon on AI agents) · TechCrunch (Yelp Check for Me)

Frequently asked questions

How do I know if my website is exposed to the October 2026 PHP flaws?

You need to know which PHP version your server runs. The patched versions are 8.5.12, 8.4.27, 8.3.36, and 8.2.35, released on October 21. You can check in your hosting control panel or ask your host's support team. If your site runs on PHP 8.1 or older, it no longer gets official updates and needs to move to a supported version.

Can I use Claude Duet in Italy?

Not yet. Claude Duet is a waitlist-only research preview, starting with Pro and Max subscribers in the United States. Anthropic hasn't given dates for other countries. In the meantime, it's worth deciding which apps you'd be willing to connect to an assistant that works on its own.

Do locked ChatGPT chats protect customer data?

Only from whoever picks up your phone or computer. Locking with a PIN or Face ID hides the conversations, but OpenAI makes clear it isn't end-to-end encryption or a separate vault: the data stays on its servers just as before. And if you reset the PIN, your locked chats are deleted for good.

Want to know if your website is up to date, and which of these updates your business actually needs?
A half-hour chat, no strings attached, in plain English, no tech jargon
Let's talk →