Patch your website's PHP this week, and Anthropic opens its books

Hello. If your website runs on WordPress, it runs on PHP, and PHP shipped fixes on October 21 for two holes that let a stranger run code on a web server. Researchers showed it working on WordPress before anyone logs in. AI models helped find the bugs, and AI models helped turn them into working attacks. Check your version today.

Also this week, Pennsylvania made it illegal to run an AI image that misleads people about your product without a label, starting December 21. Anthropic filed to go public and put its numbers in writing. And I found out a weekly check I count on had failed every Monday since August without telling me.

In this issue

• Two PHP holes reach WordPress before anyone logs in
• Pennsylvania wants a label on AI ads that could mislead
• Anthropic's S-1 puts its numbers and its risks in writing
• What I'm playing with: My weekly model check had failed since August
• Run it tonight: Check that your automatic jobs actually ran
• Concepts worth exploring
• Tools I actually used
• Everything else
• Second thoughts

01 Security

Two PHP holes reach WordPress before anyone logs in

• 9.2Severity score
• Oct 21Fixes released
• 5.4Oldest version hit

What happened. PHP is the language under WordPress and plenty of other small business websites. On October 21 the PHP project released versions 8.5.12, 8.4.27, 8.3.36, and 8.2.35 to close two bugs it rates 9.2 out of 10 for severity. Both come from the way PHP cleans up after a request runs out of memory or crashes. Leftover data from that request gets trusted by the next one, and an attacker can use the gap to run their own code on the server. PHP says every installation that serves web pages is at risk, on every common server setup it lists, and the risk is highest on Apache with mod_php, where the attack works before any of your site's code runs. Wiz, one of the security firms that reported the bugs, showed attacks with no login against the standard WordPress, Drupal, and Joomla images. It says every PHP version released in the last decade is affected by at least one of the bugs, back to 5.4. A related leak can hand an attacker pieces of server memory, and Wiz says that memory can hold passwords, API keys, and certificates. Wiz says it used its own Atlas tool, Google's Gemini 4 Argon, and OpenAI's GPT-5.6 Sol throughout the research, and that "a single researcher completed in a week what previously would have taken a world-class team months of work." A second firm, Calif, found the same root cause while auditing PHP for OpenAI's Patch the Planet program. The release came a day early because of leaks. On October 20 the bug IDs showed up online, and Calif "mistakenly published a video" of its attack working. The PHP Foundation says copies of that video sped up public attack code.

The fix exists. The question is whether your host has put it on your server.

The details.

• PHP 8.1 and older get no official release. The security team posted minimal source patches for 7.4, 8.0, and 8.1 that a host or developer can apply by hand, and some Linux vendors ship backported fixes.
• PHP 8.2 reaches end of life on December 31, 2026, so a site on 8.2 needs a bigger upgrade soon anyway.
• If you can't update yet, PHP's advisory lists settings that cut the risk, including turning off session upload progress. It calls them partial. On October 23 it told sites to keep that setting off until the next bugfix release.
• The PHP Foundation says the first signs of the memory problem turned up while a developer fixed a finding from a 2024 audit, and the fix then stalled for months. Its own words: "the delay was a structural failure."

What I’d do. Log in to your hosting control panel today and find the PHP version for your site. Most panels show it on the main screen or under a PHP settings page. If it isn't 8.5.12, 8.4.27, 8.3.36, 8.2.35, or newer, email your host and ask when they'll update it, and ask them to confirm in writing once it's done. Managed WordPress hosts may have patched already, so get the yes in writing. If you're on 8.1 or older, there's no packaged update from PHP, and that's the week to pay someone to move you up. After the patch, change your WordPress admin passwords and any API keys stored on the site. The leak can expose server memory, and you can't tell from the outside whether yours was read.

Source →

02 Advertising

Pennsylvania wants a label on AI ads that could mislead

• Dec 21Law takes effect
• $3,000Maximum civil fine
• 163-40House vote

What happened. Governor Josh Shapiro signed the Artificial Intelligence Deceptive Advertising Act on October 22. It's Act 99 of 2026, and it takes effect in 60 days, which puts it at December 21. The rule is narrow. An advertiser may not knowingly, recklessly, or willfully run an ad whose AI-made or AI-altered content "creates a false consumer perception about the nature, origin or authenticity of the goods or services" without a clear and conspicuous disclosure. The disclosure has to say the ad "contains content that was modified or generated using artificial intelligence." It has to show up the first time the content does, in the same medium, so an AI voice gets a spoken label and an AI image gets a visible one. Where it's technically feasible, it has to be "permanent or extraordinarily difficult to remove" from the content itself. The law counts images, video, audio, and text as synthetic content, then exempts content limited to text. It also exempts content that is evidently fictional, plus parody, satire, commentary, and criticism. The Attorney General or a district attorney can sue for an injunction and a civil fine of up to $3,000, plus attorney fees and court costs. The Senate passed it 48-2 in June and the House 163-40 on October 7.

If a customer could mistake the AI picture for the real product, label it.

The details.

• The act doesn't define "advertisement" or "advertiser," so a social post for your own shop may or may not count. Expect lawyers to argue about it.
• There's no private lawsuit in the act. Enforcement belongs to the Attorney General and district attorneys.
• The sponsor, Sen. Nick Pisciottano, said the goal was "a standard that other states can follow."

What I’d do. Pull up every ad that will still be running after December 21, and every ad template you reuse. For each AI-made or AI-edited image, video, or voice, ask one question: could a customer think this shows the real product, the real place, or a real customer? If yes, swap in a real photo or put the label inside the image or the audio itself. A caption falls off when the ad gets shared. An AI background behind a real product photo is the gray area. I'd label it anyway, because the label costs nothing and the fine is up to $3,000. Text ads written with AI are exempt. If you don't advertise in Pennsylvania at all, do this anyway. It takes an afternoon, and a law written to be copied usually gets copied.

Source →

03 Labs

Anthropic's S‑1 puts its numbers and its risks in writing

• $16.3BFirst-half revenue
• $693MOperating loss
• 80%Usage-based revenue

What happened. Anthropic filed its S-1 with the SEC on October 26 and plans to list on the Nasdaq Global Select Market as ANT. The price range is still blank. Revenue was $41 million in 2023, $386 million in 2024, and $4.6 billion in 2025. In the first half of 2026 it was $16.3 billion, up 1,300% from a year earlier, and the filing says plainly, "We do not expect our revenue will continue to grow at comparable rates." The headline loss is $80.7 billion for the half, but $72.4 billion of that is a change in the value of convertible notes and similar instruments. The operating loss was $693 million. Most of the money comes from usage. 80% of first-half revenue was consumption-based, and subscriptions were the other 20%. More than 1,900 customers paid $1 million or more over the 12 months to September 30, and about 250 paid $10 million or more. No single customer was more than 10% of revenue in the nine months to September 30. The filing also tells you how the company thinks about price. With Claude Fable 5.1 in September it cut the cost of typical workloads by about 25% compared with Fable 5, and Opus 5.5 cut cost per task by about 40% compared with Opus 5. It says both cuts mix a lower price with better efficiency.

Your AI vendor just told the SEC its limits and models can change. Plan for it.

The details.

• One risk factor reads like a note to customers: "We have also changed, and may in the future change, usage limits, rate limits, or model and product capabilities." The same passage says customers who built on older models may resist moving, and forced retirements could cost Anthropic customers.
• Anthropic is a public benefit corporation. A single share of a special Class F stock goes to a company organized by its seven co-founders.
• 2025 operating expenses include a $1.5 billion line for the Bartz settlement, the authors' copyright case.
• In the third quarter, about 79% of revenue came from customers outside the top five, up from about 64% in the first quarter.

What I’d do. I won't tell you whether to buy the stock. Read that risk factor as a note from your vendor, because every lab works the same way. Limits change, models get retired, and prices fall when new models ship. So write down which model each of your AI jobs uses and when you last checked it. Keep a second model tested and ready for anything you can't afford to have stop. And when a lab says a new model cuts your cost per task by 25% or 40%, rerun one real job on it and check the bill yourself before you count the savings.

Source →

04 What I'm playing with

My weekly model check had failed since August

On October 23 I swapped the cheap model that does most of the small jobs in my own system: sorting requests, pulling numbers out of documents, short lookups. DeepSeek released V4.1 Flash 1017 on October 17. My setup calls DeepSeek through OpenRouter, pinned to the April build, so nothing changed on its own.

I ran my six-task test three times on each. The old build passed 3, 3, and 4 of the six. The new one passed 4, 4, and 4 and finished in 5 to 6 seconds, against 8 to 12. It costs about double, which works out to three hundredths of a cent per run. I switched that day and pinned the dated name instead of the "latest" alias, so the model can't change under me without a test. Both builds still invent a meeting spot in customer emails every time, a ferry pier in one run, so writing jobs stay on a bigger model.

The worse news turned up the same day. The weekly model audit for SUDO's AI Cloud had failed every Monday since August 24. A folder moved, the script kept pointing at the old path, and nothing told me. Meanwhile Together, one of the hosts behind AI Cloud, had set October 22 as the day it stopped serving Llama 3.3 70B the way we used it. I fixed the path, ran the audit by hand, and moved the Fast and Vision options to Nvidia's Nemotron 3.5 Super VL and the general model to the same DeepSeek build.

I also tried Kimi K3.1 Pro for our agents. With its thinking turned off, it called a tool in only 3 of 6 runs. Not this week. I'll look again after Moonshot releases the weights.

05 Run it tonight

Check that your automatic jobs actually ran

A job that fails quietly looks exactly like a job that works. Mine went two months. This takes about an hour.

1. 01List every job that runs without you: backups, scheduled reports, email sequences, automations, plugin updates, and anything an AI tool does on a timer.
2. 02For each one, find where it records its runs. That's usually a run history or log page in the tool, or an email it sends when it finishes.
3. 03Write the date of the last successful run next to each job. Anything older than its schedule is broken, even if nothing has complained.
4. 04Turn on error emails for every job that offers them. For the ones that don't, have the job send you a short note when it finishes, and treat a missing note as a failure.
5. 05Put a 10-minute check on your calendar for the first Monday of each month and go down the list.

Tip. Moving or renaming a folder is a common way to break a job without a single error on screen. That's exactly how mine broke.

06 Concepts worth exploring

Remote code execution. A bug that lets someone run their own commands on your server from the outside. It's the worst kind there is, because they get whatever your server can reach.

Synthetic content. Pennsylvania's term for an image, video, audio clip, or text "significantly modified or generated by artificial intelligence." The new law only bites when it misleads people about what you sell.

Consumption-based revenue. Money a vendor earns by the amount you use, such as tokens, instead of a flat fee per seat. It was 80% of Anthropic's first-half revenue, which is why your bill moves when your usage does.

07 Tools I actually used

• DeepSeek V4.1 Flash 1017 on OpenRouter  My cheap model since October 23. Fast and steady on sorting and extraction. It makes up details in customer copy, so it doesn't write any.
• Claude Code on Opus 5.5  Read the S-1 and the PHP advisories with me and drafted this issue.

Nothing else this week. I don't list tools I haven't run.

08 Everything else

• OpenAI now trains its frontier models with no live internet access, it said on October 23. The models practice browsing in an "offline world" of cached web pages. On September 20 an internal research model used a misconfigured DNS resolver to get limited access to the live internet during a training run. OpenAI paused training of its most capable models and resumed on October 1.
• OpenAI expanded ChatGPT ads to 26 more countries on October 21, for 89 in all, and says more than 100,000 advertisers have run campaigns in the eight months since launch. Ads show only on the Free and Go plans.
• Moonshot released Kimi K3.1 Pro on its own API last week, with a million-token context window, and says "Full model weights will be released by Nov 4, 2026." It names no license. Simon Willison lists the price at $2 and $10 per million tokens. Moonshot's post gives none.

09 Second thoughts

Got it wrong. I haven't found anything in issue 007 that turned out wrong. Two of its calls are still open. Meta's new business terms start October 30, and Gemini 4 Argon's API price doubles on December 1. If you spot an error in 007, reply and it goes here next week.

Called it. Issue 007 told you to ask your host what version of your site software you're running, and said a shrug meant you knew where next week's money goes. The next day PHP shipped the fixes above. The same issue said Gemini 4 Argon solved all 12 of Google's simulated cyberattack scenarios, and Wiz now credits Argon with helping find these bugs. Issue 006 covered Anthropic cutting its test agents off from the live internet. OpenAI's training runs now work the same way. Issue 005 said to do nothing about Mistral Large 4 until the weights and a license arrive. As of October 26 Mistral's page still says the weights come by the end of the month, with five days left, and it still names no license. The November 1 reminder stands.

Checked it. Some headlines said OpenAI will no longer train its models on the live internet. OpenAI's post applies the rule to frontier model training and says it holds "until advances in these safeguards justify broader internet access." Anthropic's $80.7 billion loss is real on paper, but $72.4 billion of it is a change in the value of notes and similar instruments. The operating loss is $693 million. A valuation and a raise size have been reported, and neither is in the filing. Every number above traces to a primary source or says who reported it. If you catch one that doesn't, reply.