AI news · week of Oct 23, 2026
16 tools and stories, 2 security alerts, 1 sponsor
Matt Wolfe’s weekly roundup: Anthropic opens a waitlist for Claude Duet, two critical PHP flaws put every WordPress site at risk, a Railway breach exposes API keys, and big sites start blocking AI agents. Plus new tools from OpenAI, Ideogram, ElevenLabs, Perplexity, Mistral, Canva, Zoom and more. Tap any ▶ time to jump the video.
Tools in this episode
Tap a name for the full card. Badge = what you can do with it today.
- 01WAITLISTClaude DuetIn Claude Pro & Max (US)▶ 0:00
- 02PATCH NOWPHP2Shell: two critical PHP flawsUpdate PHP now▶ 3:05
- 03ROTATE KEYSRailway breach: environment variables exposedRotate every env var▶ 5:29
- 04LIVEOpenAI’s 28 days of shipping: week 3Codex & Work users▶ 8:57
- 05ROLLING OUTChatGPT Locked ChatsFree on every plan▶ 11:12
- 06APIOpenAI Decisions API (generally available)$0.10 / 1M input tokens▶ 11:36
- 07LIVEIdeogram VectorAPI $0.13–$0.32 per SVG▶ 12:25
- 08LIVE · PAIDElevenLabs ChimeraFrom Starter, $6/mo▶ 13:24
- 09OPEN WEIGHTSPerplexity pplx-decider v1.2$0.01–$0.02 / 1M input▶ 16:14
- 10LIVEPerplexity Computer runs fully local on MacPro, Max & Enterprise▶ 16:46
- 11TRENDWebsites are blocking AI agentsNot a product▶ 17:14
- 12LIVEYelp “Check for Me” restaurant callsIn the Yelp app▶ 20:28
- 13APIMistral Voxtral 3$0.0035/min transcription▶ 20:56
- 14LIVECanva CraftAll Canva users▶ 21:47
- 15RUMORRumor: OpenAI “Aurora” image modelsNot released▶ 22:30
- 16DEC 2026Zoom deepfake risk detectionEnterprise; price TBA▶ 23:12
- 17RESEARCHMicrosoft MAI-DxONot available to use▶ 23:48
- 18DEMOUnitree H2 humanoid stunt videoNo public price▶ 24:25
Security alerts
Two stories from the security segment that need action if you run a website or host apps.
PHP2Shell: two critical PHP flaws
Who: Found by researchers at Calif using OpenAI’s GPT-5.6 Sol (OpenAI’s Patch the Planet program), and independently by Wiz. Fixed by PHP’s maintainers.
Fixed in PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. Versions 8.1 and older are end-of-life.
- What it is
- Two critical bugs in PHP, the language behind WordPress, Wikipedia and Drupal. They can let an attacker run their own code on a server without logging in. They are CVE-2026-109618 (OOM corruption) and CVE-2026-109619 (stale global state).
- What’s new
- Patches shipped Oct 21. The bugs exist in roughly 20 years of PHP versions. Cloudflare’s Matthew Prince said attacks went from thousands in the 10 hours before disclosure to millions afterward. Wiz called it among the most severe vulnerabilities in software history.
- Best for
- Anyone running a PHP site, including every WordPress site.
- Ask your host which PHP version you’re on. You want 8.5.12, 8.4.27, 8.3.36 or 8.2.35 or newer.
- On PHP 8.1 or older, upgrade to a supported version. Those versions only got source backports.
- Can’t update yet? PHP’s interim fix for CVE-2026-109619 is
session.upload_progress.enabled=0in php.ini, then restart PHP. PHP’s Oct 23 follow-up also recommends it. - Update WordPress core and plugins after taking a backup.
A 20-year-old bug that was too hard to exploit became practical with AI. That shows how much havoc AI can wreak on the web, even though AI also helped find it.
Context: Matt says AI found it before attackers did. Amazon reported an exploit attempt on Oct 17, days before the patch, and the PHP Foundation says it knows of no successful remote code execution.
PHP security notice ↗ The Verge timeline ↗ Patch the Planet ↗
Railway breach: environment variables exposed
Who: Railway
Railway says commits pushed during the pause weren’t deployed. Redeploy, reconnect GitHub, and recreate CLI and CI credentials.
- What it is
- Railway, a cloud platform where developers host app backends, detected that an unauthorized third party may have accessed encrypted environment variables, and decrypted some of them.
- What’s new
- Disclosed Oct 21. Environment variables often hold API keys (for example OpenAI or Anthropic keys) and database or registry credentials. Railway believes it was part of attacks on platforms aimed at AI inference providers, and it paused all deployments while it investigated.
- Best for
- Developers and companies hosting apps on Railway.
- If you use Railway, rotate every environment variable now: API keys, database passwords and registry credentials.
- Revoke the old keys at each provider (OpenAI, Anthropic, Stripe and so on), not just in Railway.
- Check provider billing and logs for unusual usage since Oct 21.
- Redeploy your latest commit and recreate CLI and CI tokens.
Stolen API keys mean someone else runs up your AI bill. If it can happen to Railway, it could happen to Vercel or any host, and right now not much feels safe.
Tool cards, in the order covered
Claude Duet
by Anthropic
Research preview with a waitlist, starting with Pro and Max users in the US. Included at no extra cost and shares your plan’s usage limits. Claude Pro is $20 a month ($17 a month billed annually); Max starts at $100 a month. Standalone iOS and Android app, plus claude.ai.
- What it is
- A proactive personal-assistant agent app. It sends a short morning brief, watches your connected apps during the day, messages you only when something needs attention, and checks with you before it acts on your behalf.
- What’s new
- Opened as a research preview. Built around Claude Opus 5.5. Ships with a dedicated classifier and safety system prompt that checks each message and blocks sharing the strictest tier of data, such as credit card numbers and government IDs. Matt lists connectors including Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn, DocuSign, HubSpot, Canva and Shopify (as shown in the demo).
- Best for
- Professionals and small-business owners juggling email, sales, payroll and support across many apps.
Matt calls the in-chat dashboard cleaner than what he gets from ChatGPT’s Dots or Meta’s Muse, and he’s excited to try it, but he’s still on the waitlist, so there’s no hands-on demo yet.
OpenAI’s 28 days of shipping: week 3
by OpenAI
Codex and ChatGPT Work users. The daily updates are posted on X by OpenAI’s Thibault Sottiaux.
- What it is
- OpenAI’s pledge to ship one clear improvement to Codex or ChatGPT Work every day for 28 days, or give users a usage reset on days it ships nothing.
- What’s new
- As stated in the video: Day 12 stretched effective usage (context reused across subagents, longer caching, better request routing). Day 13 was a reset. Day 14 let you edit an earlier message in a thread in the desktop app; everything after the edit is discarded. Day 15 was a reset. Day 16 added 12-hour prompt caching in the API, and logged-in Codex and Work users keep cache indefinitely. Day 17 was a reset.
- Best for
- Developers and teams who use Codex or ChatGPT Work daily.
Mostly minor this week. He thinks the Day 16 caching change matters most because it means fewer tokens burned when you come back to an old conversation, and he admits the Day 14 message-editing change confused him.
ChatGPT Locked Chats
by OpenAI
Free. Rolling out gradually on ChatGPT Free, Go, Plus, Pro, Business, Enterprise and Edu. Not available on ChatGPT for Healthcare, or for temporary, archived, group or shared-link chats.
- What it is
- Lets you lock individual ChatGPT conversations behind a four-digit PIN, or Face ID or Touch ID on supported devices. Locked chats move to their own section, out of regular history and search.
- What’s new
- Announced Oct 20. Locked chats get their own separate memory, so new details in them don’t leak into regular chats. Resetting the PIN permanently deletes all locked chats. It is not end-to-end encryption and doesn’t change how long OpenAI retains data.
- Best for
- Anyone who hands their phone or laptop to family or coworkers and keeps personal chats in ChatGPT.
It isn’t in Matt’s account yet. If you don’t see it, it just hasn’t reached you.
OpenAI Decisions API (generally available)
by OpenAI
Generally available. $0.10 per million input tokens, $0.01 per million cached input tokens, and no output-token charges (OpenAI, Oct 22).
- What it is
- An API for questions with bounded answers. Instead of returning text, it returns a probability, a choice from your list, or a score against a rubric, so apps and agents can classify, route and decide quickly.
- What’s new
- Now generally available, built on GPT-6 Luna. OpenAI says it is up to 10x faster than asking GPT-6 Luna through the Responses API and accepts up to 400 questions per request. It adds native GPT-Live integration and prompt caching.
- Best for
- Developers building agents, support routing, moderation or classification into apps.
Matt notes it’s roughly the third time OpenAI has announced this API, and he sees it as OpenAI’s answer to a rival decision model that was all the rage a few weeks ago.
Ideogram Vector
by Ideogram
On ideogram.ai. API pricing is per SVG: $0.13 (low), $0.16 (medium) and $0.20 to $0.32 (high quality) for text-to-vector. In-app plan pricing wasn’t stated in the video.
- What it is
- A model that generates clean, editable SVG vector art (logos, illustrations, print and sticker designs) from a prompt, or converts an existing image into a vector.
- What’s new
- A brand-new vector model. Because the output is true SVG, it scales to any size and you can select and recolor individual shapes.
- Best for
- Designers, merch and sticker sellers, and small businesses that need logos or print-ready art.
He had a few days of early access. His wolf-and-moon vector came out well, but of four monkeys-on-roller-skates results, two had extra legs. Vector is not as dialed in as Ideogram’s image generation yet.
ElevenLabs Chimera
by ElevenLabs (with ByteDance’s Seedance 2.5 video model)
Paid. On ElevenLabs’ pricing page, Image & Video is included from the Starter plan ($6 a month) up, not on the Free plan. No separate Chimera price is published. Matt hit a 30-second limit on the reference audio.
- What it is
- A workflow in ElevenCreative that makes a character talk on camera. You upload or create a character, write the line, pick or design a voice, and direct the delivery with audio tags like [whispers] and [laughs].
- What’s new
- Launched Oct 21. It pairs ElevenLabs’ Eleven v4 voice model with Seedance 2.5 for the face and lip sync. It’s a template that chains existing models, not a new model.
- Best for
- Creators making talking-head clips, explainers and character videos.
The lip sync is decent and getting closer, but it ignored his photo and invented its own shots, including putting him on a stool. The UI also fought him: the submit button scrolled off-screen, and the 30-second limit error appeared only after he’d generated the speech.
Perplexity pplx-decider v1.2
by Perplexity
Open weights, or the hosted Decisions API at $0.02 (27B) and $0.01 (4B Flash) per million input tokens. Output tokens are free (per Perplexity’s API changelog).
- What it is
- Perplexity’s own decision models: given text or an image, they return probabilities for yes/no, multiple-choice or rubric questions instead of writing prose.
- What’s new
- A 27B model (pplx-decider-v1.2) and a new 4B Flash model, released with open weights under Apache 2.0. Version 1.2 answers several questions in one pass and is up to 5x faster.
- Best for
- Developers who want cheap classification or routing they can self-host.
He read Perplexity’s own chart as putting the 4B model on top. Independent checks of the public leaderboard rank the 27B model well above it, and the 4B as the best of the 4B-size models.
Perplexity API changelog ↗ Aravind Srinivas announcement (X) ↗
Perplexity Computer runs fully local on Mac
by Perplexity
Pro, Max and Enterprise subscribers, in the Perplexity Mac app on supported Apple silicon Macs. 36GB of memory or more is recommended for fully local mode. Hybrid Compute (local plus cloud models) works on Macs with as little as 24GB.
- What it is
- Perplexity’s agent that plans multi-step tasks and uses tools and apps. It can now run entirely on your Mac: the model, the agent and the task execution.
- What’s new
- Launched Oct 20. It runs Qwen 3.6 or PPLX Pro (Perplexity’s post-trained Qwen) on-device. With permission it works across Notes, Messages, Mail, Keynote and Numbers without sending personal data to the cloud, and you approve cloud access (and credit spend) when a task needs the web.
- Best for
- Mac users who want an AI agent working on private files, like finances or client documents, that never leave the machine.
He covers it on the way to the bigger theme of the week: who will let AI agents in and who will lock them out.
Websites are blocking AI agents
by Industry-wide (reported by The Verge)
Not a product. It affects which sites your AI agent (Dots, Muse, Grok, Claude Duet) can actually use.
- What it is
- Big services are shutting AI agents out. Yelp only allows agents that pay for its data licensing, Delta blocks agent bookings, and JPMorgan Chase and Charles Schwab block agents from accounts. Users report blocks on Cars.com and Amtrak too.
- What’s new
- Shopify has gone the other way and opened its storefronts to AI agents. Amazon was already fighting off shopping agents.
- Best for
- Anyone who uses or plans to use an AI agent to shop, book or bank, and any business that sells online.
Companies must choose between human-first and agent-friendly. Blocking agents protects big sites from becoming interchangeable, but he expects most will eventually let agents in or lose the sales.
Yelp “Check for Me” restaurant calls
by Yelp
In the Yelp app. No consumer price was stated. Yelp’s separate Smart Lead Reply for businesses costs $30 a month.
- What it is
- An AI voice agent in the Yelp app that calls a restaurant for you to check the wait time, then texts you the answer.
- What’s new
- It appears on restaurant pages. You enter your party size and phone number. The Yelp AI assistant also gained OpenTable booking, memory of your preferences, 30-day chat history and custom reminders.
- Best for
- Diners who hate calling ahead, and restaurant owners who should expect AI calls.
The irony: Yelp blocks outside agents from using its site, while it tries to become your restaurant agent itself.
Mistral Voxtral 3
by Mistral AI
Paid API, hosted in the EU, or self-deploy on your own infrastructure. Listed prices include $0.0035 per minute for Mini transcription and $13 per million characters for Voxtral Mini TTS 3.0.
- What it is
- A family of speech-to-text and text-to-speech models plus a voice-agent SDK. Listening, reasoning, tool calls and speaking run in one real-time pipeline.
- What’s new
- It transcribes 50+ languages, tells speakers apart in real time (diarization), handles interruptions, and speaks with emotion. In the demo it tracks two people talking to one agent.
- Best for
- Teams building call-center agents, voice assistants and hands-free copilots.
A really impressive voice agent. It knew which of two people was speaking, handled being interrupted and kept following the conversation.
Canva Craft
by Canva (with researchers from its Leonardo.ai acquisition)
Available to all Canva users from the Dream Lab tab, and through the Canva API. Canva didn’t state a separate price.
- What it is
- Canva’s first in-house image-generation model, trained on human-designed Canva templates rather than photos, so it handles layouts and legible text.
- What’s new
- A two-stage approach: it plans the layout first, then renders up to 2K resolution in 30 aspect ratios. Output can be turned into layered, editable designs. Faster Craft-fast and Craft-ultrafast versions are coming.
- Best for
- Canva users making social posts, posters, menus and presentations.
Matt discloses a small investment in Canva and says it doesn’t affect his coverage.
Rumor: OpenAI “Aurora” image models
by OpenAI (not announced)
Not released. Rumor only.
- What it is
- Three model names (aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha) briefly appeared in OpenAI’s API and then in blind arena image tests.
- What’s new
- TestingCatalog and others think it is GPT-Image-3. OpenAI hasn’t confirmed anything.
- Best for
- Anyone choosing an image tool right now.
It isn’t confirmation, but leaks like this usually come shortly before a launch. He won’t be surprised by a new OpenAI image model within a couple of weeks.
Zoom deepfake risk detection
by Zoom
Expected December 2026, for enterprise admins. Pricing not announced.
- What it is
- Real-time analysis of video and audio in Zoom calls that flags participants who may not be who they claim to be, with a confidence indicator.
- What’s new
- Announced at Zoomtopia on Oct 20, alongside centralized AI management for admins, custom AI guardrails and sensitivity-label integration.
- Best for
- Companies hiring remotely or approving payments over video calls.
He points to the deepfake-CEO scams, where staff wired money because a fake boss on a live video call asked them to.
Microsoft MAI-DxO
by Microsoft AI
Research only. It isn’t available to patients or doctors as a product.
- What it is
- Microsoft AI’s diagnostic system, which works through a medical case step by step, choosing tests and reaching a diagnosis.
- What’s new
- Microsoft reports 80% accuracy vs. 47% for physicians on emergency-department cases (about 1.7x). On complex NEJM cases it reports more than 2x the accuracy of physicians using textbooks and the internet, and nearly 4x without them.
- Best for
- Clinicians and anyone following AI in healthcare.
Some version of this may end up in your doctor’s office, at least to double-check diagnoses.
Unitree H2 humanoid stunt video
by Unitree Robotics
Unitree doesn’t list a public price for the H2 on its product page.
- What it is
- A viral promo of Unitree’s H2 humanoid robot doing flips over a car, kicking a hood, jumping from 4 meters, lifting a car off a trapped person, and shoving a kid out of a car’s path.
- What’s new
- A demo video, not a product launch. The H2 is listed on Unitree’s site.
- Best for
- Robotics fans, and skeptics of staged demos.
He doubts any of it is autonomous and thinks someone is driving it. He doesn’t believe it really lifted the car by a plastic bumper, and says the push that saves the kid looks more dangerous than the slow-motion car.
Chapters
- ▶ 0:00 Claude Duet: Anthropic’s new AI agent
- ▶ 2:47 Major AI cybersecurity threats (PHP2Shell, Railway)
- ▶ 7:54 Hyperagent (sponsor)
- ▶ 8:57 OpenAI’s latest updates
- ▶ 12:25 Ideogram’s new AI vector generator
- ▶ 13:24 ElevenLabs Chimera
- ▶ 16:02 Perplexity’s new AI models and Computer
- ▶ 17:14 Websites are blocking AI agents
- ▶ 20:28 Yelp’s AI agent and Mistral Voxtral 3
- ▶ 21:47 Canva Craft and new AI image models
- ▶ 23:18 AI deepfake detection and medical AI
- ▶ 24:29 Unitree’s humanoid robot demo
Show more
Want this section to go deeper? Share this page with the creators and have them sign up so we can get permission to dig deeper into this episode.
People
On this episode
Matt Wolfe
Official site · YouTube · X · Newsletter
Host. Runs FutureTools and records this roundup on Thursdays. ▶ 0:00
People mentioned
Matthew Prince
Cloudflare CEO. Reported thousands of PHP exploit attempts before disclosure and millions after. ▶ 4:20
Thibault Sottiaux
OpenAI. Posts the daily “28 days of shipping” updates for Codex and Work. ▶ 9:59
Show more
Want this section to go deeper? Share this page with the creators and have them sign up so we can get permission to dig deeper into this episode.
Products
Macs that meet Perplexity’s memory guidance for running Computer on your Mac.
Apple 2026 MacBook Pro 14-inch (M5 Pro, 24GB, 1TB) — Meets the 24GB minimum for Perplexity’s Hybrid Compute mode. Amazon ▶ 16:46
Mac Studio (M4 Max, 36GB, 512GB) — A desktop that meets the 36GB recommendation for running Perplexity Computer fully local. Amazon ▶ 16:46
Apple 2024 MacBook Pro 16-inch (M4 Max, 36GB, 1TB) — A laptop that meets the same 36GB recommendation. Amazon ▶ 16:46
Apple 2026 MacBook Pro 16-inch (M5 Max, 36GB, 2TB) — The newer M5 Max version with 36GB of memory. Amazon ▶ 16:46
Apple Mac mini (M5 Pro, 48GB, 1TB), configured to order — Perplexity highlights the new Mac mini; this 48GB build clears the 36GB recommendation. Amazon ▶ 16:46
Show more
Want this section to go deeper? Share this page with the creators and have them sign up so we can get permission to dig deeper into this episode.
Discussed on the show
Websites blocking AI agents — Yelp, Delta, Chase and Schwab shut agents out while Shopify opens up. ▶ 17:14
OpenAI “Aurora” image model leak — Three model names surfaced in the API and blind arena tests. ▶ 22:30
Deepfake CEO scams — Why Zoom is adding deepfake risk detection. ▶ 23:12
Unitree H2 stunt video — Matt doubts the stunts are autonomous. ▶ 24:25
Show more
Want this section to go deeper? Share this page with the creators and have them sign up so we can get permission to dig deeper into this episode.
Places
None this episode.
Show more
Want this section to go deeper? Share this page with the creators and have them sign up so we can get permission to dig deeper into this episode.
Sponsor
Hyperagent ▶ 7:54 — A platform for “hiring” AI agents that work like employees (research, content, analysis and operations), with multiple agents and people in the same room. In the sponsored read, Matt says you can sign in with your existing ChatGPT plan instead of paying for another AI subscription, or pick a paid plan and get $100 in bonus credits.
Show more
Want this section to go deeper? Share this page with the creators and have them sign up so we can get permission to dig deeper into this episode.
Show notes are a summary, not a full transcript. Prices and availability are as announced by each company on the dates noted and can change. As an Amazon Associate we earn from qualifying purchases. Links on this page may be affiliate links; if you buy through them, we may earn a commission at no extra cost to you. The Hyperagent link is the creator’s sponsor link.
More AI news notes: FutureTools on My Show Notes