AP HOLDINGS
EN
Menu
← AP News
NO. 27 · Weekly News

Weekly AI News — OpenAI's Models Move Into Google Cloud, AI Found a Big Hole in PHP First, and ChatGPT Ads Now Run in 89 Countries

Weekly AI News · OpenAI · Google Cloud · PHP2Shell · Security Patches · ChatGPT Ads · 2026.10.24 · AP Holdings Editorial · 10 min read

Translated from the Korean original by AI; native-speaker review pending.

Three major announcements landed on the same day, October 21. OpenAI agreed to sell its models on Google Cloud too, two critical PHP flaws found with AI were patched in an emergency release, and ChatGPT ads added 26 countries to reach 89. For each story we lay out what is fact, how far it has been confirmed, and what a small team can check this week.

This article was drafted with AI (Claude) and reviewed by an editor. Anthropic, which makes Claude, competes with OpenAI and Google, both of which appear in this article, and our company uses Claude at work, so we have a conflict of interest. Figures announced by companies are labeled "announced by the company," and where reports differ, we give both sides. Nothing here is investment advice about any company.

As it happens, this week's big stories all came out on October 21. One is about where you buy (OpenAI's models can now be bought on Google Cloud as well), one is about who finds it first (AI found critical flaws in PHP, a language much of the web runs on, before anyone else), and one is about who pays (ads on the screens of free ChatGPT users expanded to 89 countries). The donation of Agent Skills to a foundation, the new Claude API version and Railway's key-leak notice, all from the same week, were covered separately in NO.26.

1. OpenAI's Models Are Coming to Google Cloud (Fact)

On October 21, OpenAI announced that its top models and its coding agent Codex will be available in Model Garden, Google Cloud's model catalog service. According to the announcement, purchases go through Google Cloud Marketplace, and companies that already have spending commitments with Google Cloud can draw down those commitments. General availability is planned "in the coming weeks," and the models will also come "soon" to Google's coding tool Antigravity and its business automation agent Gemini Agent. The announcement added that Microsoft remains OpenAI's "primary cloud partner."

OpenAI's announcement itself names no models and gives no prices. SiliconANGLE and Techzine reported that three models are included, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, and Techzine wrote that it starts as a private preview for select enterprise customers. "General availability within weeks" (OpenAI) and "starting as a private preview" (Techzine) do not contradict each other, but they mean it is not yet open to everyone right now.

2. AI Found Critical Flaws in PHP First — "PHP2Shell" (Fact)

The same day, OpenAI introduced its open-source security program "Patch the Planet." It brings together researchers, maintainers and incident-response bodies to fix critical flaws in open-source software, and its first case is two flaws in PHP, the language that runs websites and web apps. OpenAI said a researcher at the security firm Calif found one flaw with GPT-5.6 Sol within this program, and that Wiz independently found and reported both flaws. TNW and Latent Space reported that the first discovery was in August, and that Wiz used its own tool together with Google's Gemini 4 Argon and GPT-5.6 Sol (Wiz is a company acquired by Google). OpenAI said the parties coordinated for about two months before disclosure.

The flaws are CVE-2026-109618 and CVE-2026-109619. According to the alert from the Canadian Centre for Cyber Security, they let an attacker run arbitrary code on a server without authentication, and the fixed versions are PHP 8.5.12 · 8.4.27 · 8.3.36 · 8.2.35 (released October 21). Every PHP installation that serves web requests is in scope, and the alert says Apache's mod_php setup is the most at risk. Versions 8.1 and older are no longer supported by the PHP project, so anyone still using them is told to check directly whether their vendor (hosting company or device maker) has a fix.

One number needs to be read correctly here. The "70% of the web" in many headlines is an estimate of the share of sites that run on PHP (Wiz), not a count of sites that are actually vulnerable (TNW). The state of real attacks also differs by source. Reports agree that attack code has been circulating since disclosure, but The Hacker News headlined that attack attempts number in the millions, while the Canadian Centre for Cyber Security wrote that, as of its October 23 update, no active exploitation had been confirmed.

On October 23, a separate flaw, CVE-2026-115560, was announced. The October 21 patches do not fix it, but it applies only to a rare combination in which session settings are changed at runtime, so PHP said it will fix it in the next regular release rather than with an emergency patch. Until then the temporary measure is `session.upload_progress.enabled=0`, and THN and the Canadian Centre for Cyber Security both said that sites that already applied this setting on the 21st need no further action.

OpenAI also released a Codex skill for checking PHP versions, and said it would support organizations short of security staff with $1 billion worth of access to its security tool Daybreak (announced by the company).

3. ChatGPT Ads Add 26 Countries to Reach 89 (Fact)

Also on October 21, OpenAI announced it had opened ChatGPT ads in 26 more countries, for 89 in total. According to the announcement's title, more than 100,000 advertisers have run ads so far (announced by the company; a cumulative figure). According to Investing.com and PPC Land, 14 of the 26 new countries are in Latin America, and the ones named are Argentina, Chile, Colombia, Peru, Nigeria, South Africa, Serbia, Ukraine, Cambodia, Bangladesh and Pakistan. Advertisers can buy through agency partners or directly in Ads Manager.

Ads appear only on the Free and Go plans, not on Plus, Pro or Enterprise. South Korea is not one of these 26; ads were already introduced there on June 19 (Electronic Times, Etoday). According to Korean reports at the time, in Korea too only adult users on the Free and Go plans are targeted; no ads appear for accounts confirmed or predicted to belong to minors; and advertisers receive only aggregate performance such as views and clicks, not conversation content. OpenAI again repeated its principle that ads do not influence answers and are displayed separately from them.

Reading the Three Stories Together (Interpretation)

What the three stories have in common is that the way AI meets the things underneath it is changing. Models now go to the clouds customers already use instead of staying on one company's cloud; AI finds holes in foundational code that people have used for decades before people do; and the free AI screen has met an old money flow called advertising.

Three questions remain for a small team. First, if model choice is freed from contracts, what do we choose by? The more you can pick models from several companies inside the same cloud, the more a standard like the "30-minute test on three of your own tasks" from NO.17 matters. Second, AI finding holes first means attackers use the same tools. This time the finders and fixers were two months ahead, but attack code circulated right after disclosure. The shorter the time until a patch, the better off a team is if it already knows "where in our services that component lives." Third, if ads do not influence answers, there are two paths by which free users see our product: the ad slot and mentions within answers. OpenAI's principle is that the answer side cannot be bought with money, so we think accurate, open public pages are the foundation of that path (this last connection is our interpretation).

What We Would Do — Checks to Run This Week (Practice)

  1. Make a list of where PHP is hiding. Our homepage is a static site made with a build tool, so it is not set up to run PHP directly on a server. Even so, PHP may be inside services that others run, such as shopping-mall solutions, blog or booking plugins, and hosting admin screens. For each outside service we use, we check the notices for "whether it uses PHP, and if so, whether it has upgraded to the October 21 fixed versions," and ask if there is no notice. Filling in the boxes on our list is faster than being alarmed by the 70% figure.
  2. Test models by the task. If OpenAI, Google and Anthropic models can be picked side by side inside a cloud, the standard is "how many times did we have to fix it on our work" rather than "which company." When the new models open for general use, we will rerun NO.17's test sheet as is (three tasks · the same instruction · accuracy / number of fixes / time). Performance claims at the preview stage we only record.
  3. Pages that fit answers, rather than ads. Our promotion currently follows a principle of spending nothing on ads. So we treat ChatGPT ads as "something to watch," not something to buy now, and instead check whether our app and service pages have sentences that answer questions directly (what it does, what it does not do, sources). We could not confirm this time whether self-serve ad buying is open in Korea, so we do not assume it is.

What We Did Not Pick This Week

There was other news too: a new member of Anthropic's Long-Term Benefit Trust (LTBT) and its call for guest investigators in the life sciences, Beam, a 501B-scale open model posted on Hugging Face, OpenAI's Decisions API announcement, and DeepSeek V4.1 Flash listed on a price-comparison site. We left them out of this weekly roundup because we could not check the full text well enough or there was only one source.

What We Could Not Confirm in These Stories

  • OpenAI–Google Cloud: OpenAI's announcement has no model names, prices or regions. We confirmed the three model names only through news reports (SiliconANGLE, Techzine), and could not confirm the general-availability date beyond "within weeks." We could not open an official post from the Google Cloud side directly.
  • PHP2Shell: There is a report (TNW, citing the WSJ) that the disclosure schedule was moved up because of a leak, but we could not confirm what leaked or when. The scale of actual exploitation differs by source (see above). There are also reports that fixes for old versions 8.1 and below came from the PHP project, but the wording differs from the Canadian alert, so we wrote that vendors need to be checked. We could not open the original PHP security advisory directly.
  • ChatGPT ads: OpenAI's original announcement would not open on our side, so we confirmed it from its title and news reports. We could not confirm the 15 new countries beyond the 11 that were named, or whether self-serve ad buying is possible in Korea. We gave Korea's launch date as June 19 from Korean reports; some overseas outlets gave a different date.

Further Reading

We opened these sources of different kinds directly to check them. They are grouped into company announcements, a government agency alert and news reports.

Company announcements:

Government agency:

News reports:

← Skills became a "shared standa…AI Is Starting to Rewrite Your… →

Keep reading