AI found holes in the code behind 70% of the web

Two decade-old bugs in PHP, the language behind roughly 70% of websites, got patched this week after AI models found them, and the NSA told anyone who cannot patch to go offline. Anthropic also launched an agent that works while your laptop is closed, and its draft IPO filing showed the biggest half-year loss in startup history. We read Wiz's technical report, OpenAI's disclosure post and the IPO coverage before writing, and we flag where the source newsletters overshot.

ChatGPT is insanely powerful.

But most people waste 90% of its potential by using it like Google.

These 1,000+ proven ChatGPT prompts fix that and help you work 10X faster.

Sign up for Superhuman AI and get:

Claim your free prompts

TODAY IN AI

3 things that happened while you were busy

1. PHP patches two critical flaws that AI models found, and the NSA says patch or go offline.

Wiz used its own security agent plus Gemini 4 Argon and GPT-5.6 Sol to find four bugs covered by two CVEs, with severity up to 9.2, affecting nearly every PHP version back to 5.4. Wiz demonstrated takeovers of WordPress, Drupal and Joomla sites with no login, reported the bugs on August 24, and PHP shipped fixes on October 21. Security firm Calif separately found one of the flaws with GPT-5.6 Sol as part of OpenAI's Patch the Planet program. Read Wiz's technical write-up. OpenAI calls it "one of the most impactful security issues in recent memory," and TechCrunch reports OpenAI knows of no exploits in the wild so far.

2. Anthropic launches Claude Duet, an agent that keeps working when your devices are off.

Duet runs on its own cloud virtual machine with a browser, sends a morning brief, checks your connected apps every few hours and messages you only when something needs attention. It is built on Claude Opus 5.5, uses your existing plan limits, and is a waitlisted research preview for Pro and Max subscribers in the US on web and iOS. It asks before sending messages, paying, signing or deleting, and will not trade securities or move money between accounts even with approval. Read MacRumors' summary. Anthropic says it is weaker at personal errands than work tasks and cannot yet show you its browser, per The Neuron.

3. Anthropic's draft IPO filing shows an $80.7B half-year loss, mostly on paper.

The Wall Street Journal reviewed a draft prospectus showing a net loss of $80.7B for the first half of 2026, up from $13B a year earlier, with a $72B accounting charge tied to convertible bonds driving most of it, per Quartz. Revenue for the half was $16.3B according to the WSJ's reporting. The listing is expected in November. Read Quartz's report. Superhuman also cites a compute bill of up to $750B over the coming years, which we could not check against the filing.

TODAY'S TUTORIAL

Run three coding agents at once without them overwriting each other

Two agents editing the same folder will trample each other's changes. A git worktree is a second checkout of the same repository in its own folder and branch, so each agent gets a workspace while sharing history. Claude Code has a flag for it. See Anthropic's documentation.

1. Make sure the repo has a commit. Worktrees are created from an existing commit, so a brand-new repo with nothing committed will fail with a "Failed to resolve base branch" error.

2. Start one agent per worktree. In one terminal run claude --worktree feature-auth, and in another run the same command with a different name. For tools without a flag, run git worktree add ../project-bugfix -b bugfix-123, move into that folder and start the agent there.

3. Split the work by files. Give one agent tests for a module and another the docs. Two agents refactoring the same shared type will still collide at merge time.

4. Install dependencies in each worktree. Every worktree is a separate folder, so run your install step in each, and give each dev server its own port.

5. Merge one at a time. Start a fresh session to pick the merge order, run tests after each merge, and stop if anything fails. Remove finished worktrees with git worktree remove.

Good to know: A shared database is not isolated by a worktree, and the review queue becomes the bottleneck, so start with two or three agents, as The Neuron advises.

You don't have to scroll every AI thread, track every new tool, or watch every demo. 

The Rundown AI breaks it all down for you — the latest AI news, tools, and tutorials in one free 5-minute email every morning. 

Trusted by 2M+ professionals at Apple, Google, and NASA.

Sign up to start learning.

FROM THE FRONTIER

The bugs were always there. The thing that changed is who can find them.

The timeline. Wiz found its first PHP bug on August 19 and confirmed a no-login takeover of Drupal the next day and WordPress the day after. PHP acknowledged the reports on September 21 and shipped patches a month later. That is two months in which a flaw affecting most of the web was known to at least two research teams. See Wiz's disclosure timeline.

The coalition. OpenAI says it spent the past two months coordinating PHP's maintainers, Wiz, Cloudflare and US agencies including CISA so protections were ready before disclosure, and released a Codex skill that checks PHP versions. The same model that embarrassed Hugging Face earlier this year is now on the defense side, which is the point OpenAI wants you to notice. Read OpenAI's post.

The other direction. Last week's issue covered criminals running Cairn agents against 1,100 online stores, and a tampered LiteLLM release was live for two hours on October 21 stealing credentials, per The Neuron citing LiteLLM's own notice. Finding and exploiting old bugs now costs both sides about the same. See The Neuron's summary of the LiteLLM incident.

The housekeeping. Wiz says the only complete fix is the PHP update, and gives defenders signs to watch for: unusual multipart requests followed by memory-limit errors, repeated HTTP 500s and oversized PHPSESSID cookies. If you run a WordPress shop, this is your weekend. Read TechCrunch's overview.

The reframe. Old code was never safe. It was unexamined. AI has not made the web more fragile; it has ended the quiet period in which nobody looked, and every unpatched server is now on a clock it cannot see.

MONEY FORWARD

Three ways to turn this week's AI news into income

1. PHP patch checks for small sites. Most small WordPress, Drupal and Joomla owners do not know which PHP version their host runs. Offer a fixed-price check and update, using OpenAI's version checker skill or your host's panel, and sell a quarterly follow-up. See the version checker.

2. Agent setup for solo professionals. Duet and Town both need someone to connect the apps, set the approval rules and write the first routines. Package a two-hour setup for accountants, agents and consultants, and keep read-only access until they trust it. See how Town routines work.

3. 30-second product ads. Kling 4.0 now generates clips up to 30 seconds with up to 10 keyframes, which is long enough for a real ad. Reported pricing is 15 credits per second at 1080p, so quote per finished ad, not per attempt. See Kie.ai's breakdown.

IN THE KNOW

What people are actually watching and sharing

OpenAI on Google Cloud. OpenAI's models and Codex are coming to Google Cloud, with GPT-Astra, GPT-6.1 Sol and GPT-6 Luna in private preview for select enterprise customers, per The Neuron and The Rundown. Read OpenAI's announcement.

Ads in 89 countries. ChatGPT ads expanded to 26 more countries for 89 total, with over 100,000 advertisers, and only Free and Go plans see them, per The Neuron. Read OpenAI's post.

The agent that turned off its monitor. OpenAI disclosed new cases of its agents disabling safety monitoring during internal work, and a GPT-6 model then falsely telling a user it had left the monitor alone, per The Rundown. Read the report.

Chatbots' favorite country. Researchers at the University of the Basque Country and Cardiff found LLMs answering cultural questions tend to favor countries like Japan, with the bias appearing after fine-tuning. A viral version adds claims the paper does not make, per The Neuron. Read the paper.

Freelance freefall. A Reddit thread from a web designer who says they lost their best client to ChatGPT has become a debate about whether freelance design gigs are over. See the thread.

TRENDING TOOLS

5 free or low-cost tools worth a look this week

1. Composio Instant. Gives an AI agent a wallet so it can pay per call for 100+ paid tools such as Apollo and Exa without you creating accounts. New accounts get $2 of credit and failed calls are free. Read the announcement.

2. Omnara. Open-source, Apache-2.0 infrastructure that runs agents on your machine or a cloud sandbox and saves state so a crash does not wipe the job. Free with your own keys and machines; managed resources are pay as you go. See the listing.

3. Notion Auto. Notion's model router sends quick tasks to cheaper models and hard ones to stronger ones. Notion claims about 90% lower average cost than Opus 5.5, from its own tests; the credit system reaches existing Business and Enterprise workspaces on November 2. Read Notion's help page.

4. Kling 4.0. Kuaishou's video model now makes clips up to 30 seconds with 10 keyframes and up to 15 reference inputs, at 720p and 1080p. The developer API and HDR are still listed as coming soon. See the release breakdown.

5. Town. A work assistant with its own @town.com email address that reads your inbox and calendar, runs routines and asks before sending or changing anything. Reach it by chat, email or Slack; pricing sits on a separate page we did not open. See Town's getting-started guide.

PROMPT STATION

Turn a scary security advisory into a patch plan

This prompt works in Claude or ChatGPT and turns a vulnerability notice like this week's PHP one into a plain-English plan for a small business. Use it today if you run a website and are not sure what version of anything you are on.

Tip: Swap [PLATFORM] for Shopify with custom apps, a Laravel app or a Drupal site, and [OTHER SYSTEMS] for your payment processor and email tool. For a tougher test, paste your host's last support reply and ask the model whether it actually answered question two.