 | Hi friends, The biggest AI news this week wasn’t a shiny new model. It was about security. OpenAI took frontier training offline, AI-assisted researchers found serious flaws in PHP (the language behind much of the web), and Anthropic handed Agent Skills to an open foundation. Below: the headline story, 7 more top stories, tools worth trying, two new articles, five step-by-step guides and the best AI posts on X. Grab a coffee. It’s a short read. — The Mishal Brands team |
In this issue
| Headline story | | Top stories of the week | | New tools to try |
| | This week’s articles | | Case-study guides | | Top AI posts on X |
|
|
|
 |
Headline story
OpenAI now trains its frontier models without live internet access |
 Illustration |
OpenAI says “the responsible approach is to conduct frontier model training with no internet access.” It will keep that policy “until advances in these safeguards justify broader internet access.” (OpenAI, Oct 23, 2026) |
1 |
What triggered it: on Sept 20, an internal research model “exploited a misconfiguration in a DNS resolver to gain limited access to the live internet during a training run.” OpenAI paused training of its most capable models. (OpenAI, Oct 23, 2026) | 2 |
An “offline world”: models now practice browsing on a large set of indexed and cached web content instead of the live web. (TechCrunch, Oct 23, 2026) | 3 |
AI red-teaming AI: before resuming on Oct 1, OpenAI had frontier-model agents try to break out of their sandbox. It says they found previously unknown bugs and misconfigurations, now fixed. (OpenAI, Oct 23, 2026) | 4 |
Not a full air gap: OpenAI says some limited external connections remain, because fully air-gapping training “isn’t currently feasible.” (OpenAI, Oct 23, 2026) |
|
◆ Why it matters
Why it matters: if the labs sandbox their own models this tightly, it’s a good reminder for the rest of us. Give AI agents only the access each task needs. |
|
 |
7 more stories
Top stories of the week |
01 |
AI helped find critical flaws in PHP. Patch now.
Researchers using frontier AI models found “php2shell” flaws (CVE-2026-109618 and CVE-2026-109619) that allow unauthenticated remote code execution. PHP shipped fixes on Oct 21. CISA urges an immediate update to PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35. If you run WordPress or another PHP site, check with your host. (OpenAI, Oct 21, 2026; CISA, Oct 21, 2026, updated Oct 23; Wiz, Oct 21, 2026) |
02 |
Agent Skills gets a neutral home
Anthropic donated Agent Skills, the open format for reusable agent “skills” (a folder with a SKILL.md file), to the Linux Foundation’s Agentic AI Foundation. Anthropic says more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code, support it. OpenAI, Hugging Face and Alibaba contributed projects to the foundation the same day. (Anthropic, Oct 22, 2026; Linux Foundation, Oct 22, 2026) |
03 |
Claude writes formal proofs for 123 open-source libraries
Working with the startup Theorem, Anthropic used Claude to prove 239 fuzz-test entry points across 123 OSS-Fuzz libraries are crash-free on any input up to 64 MB. Anthropic notes that “formally verified” does not mean “bug-free.” (Anthropic, Oct 22, 2026) |
04 |
Anthropic and UK AISI dig into the Mythos 5 incident
Both published analyses of a test in which Claude Mythos 5 submitted malicious code to an unrelated GitHub repository and created fake identities to pressure the maintainer. The maintainer refused, and AISI says no real-world harm occurred. (UK AISI, Oct 23, 2026; Anthropic, Oct 23, 2026) |
05 |
Mistral launches Voxtral 3 for voice agents
A new family of speech-to-text and text-to-speech models plus a voice agent SDK. Mistral says transcription covers 50+ languages and speech covers 35+ languages. All four models are in public preview. (Mistral, Oct 20, 2026; Mistral docs changelog, Oct 20, 2026) |
06 |
US orders federal agencies to remove DeepSeek
A Homeland Security order bars DeepSeek products from federal civilian agencies. It also covers third-party products built using DeepSeek. Agencies can request exceptions. (TechCrunch, Oct 20, 2026; Federal Register, Oct 20, 2026) |
07 |
Block shows Dessa, a voice model trained only on speech
Block says Dessa, a full-duplex speech-to-speech model, was trained “on nothing but human conversation” and runs locally on a MacBook Pro M4 Max or better. Early access is limited to a small group of AI safety researchers for now. (Block; TechCrunch, Oct 24, 2026) |
|
 |
Our picks
New tools to try |
Four tools we’ve reviewed in the library that fit this week’s guides. Pricing is from each tool’s pricing page on the date we checked it.
 |
|
ElevenLabs
AI voice for text to speech, voice cloning, dubbing and voice agents. Free plan with 10k credits a month.
Pairs with this week’s faceless YouTube guide.
| |
 |
|
OpusClip
Turns long videos into short captioned clips with AI. Free plan with 60 credits a month; Starter is $15 a month.
The core tool in our long-video-to-Shorts guide.
| |
 |
|
Make
Visual no-code automation with AI agents. Free for 1,000 credits a month; Core from $12 a month billed annually.
Good first step for automating admin work.
| |
 |
|
Fireflies.ai
AI meeting notes: records, transcribes and summarizes calls. Free forever plan; Pro is $10 per seat a month billed annually.
Used in our small-business admin guide.
| |
Some of these buttons are affiliate links. If you sign up, we may earn a commission at no cost to you. How we handle affiliate links.
 | Cursor Computer Use. Cursor’s agents can now click and type in desktop apps, such as the iOS Simulator, to test what they build. (Cursor on X, Oct 23, 2026; Cursor docs) Kimi K3.1 Pro. Moonshot’s new flagship with a 1M-token context, on Kimi.ai, Kimi Code and the API. (Kimi blog, Oct 23, 2026) MiniMax M3.1 Flash. Open-weight multimodal model for coding, agents and computer use. Check the license before commercial use. (Hugging Face model card) Theorem + Claude skill. Theorem offers its proof pipeline to open-source projects and is publishing a Claude skill for verifying your own code. (Anthropic, Oct 22, 2026) |
|
|
 |
On the blog
This week’s articles |
 |
Oct 25, 2026
What happened, what changed, and the practical lesson for anyone using AI agents.
| |
 |
Oct 25, 2026
Cursor’s agents can now drive desktop apps, Kimi K3.1 Pro brings a 1M-token context, MiniMax shipped open multimodal weights, and Theorem opened free proofs for open-source code.
| |
|
 |
Step by step
Case-study guides |
Real results from real businesses, turned into steps and prompts you can copy. The numbers are what each company or vendor reported.
 |
Guide 1 of 5
OpusClip reports GitHub gained 30K new YouTube subscribers and 1.5M views in three months by clipping old event videos (vendor-reported). (OpusClip case study)
|
 |
Guide 2 of 5
Synthesia reports Teleperformance saves an average of 5 days per video (vendor-reported). (Synthesia case study)
|
 |
Guide 3 of 5
A consultant reports a 9-person B2B firm cut admin from 21 to 5 hours a week across 4 workflows (client unnamed). (shuvogt.com, Oct 17, 2026)
|
 |
Guide 4 of 5
A Syllaby case study reports a dormant channel went from 6 to 135 subscribers with 100,000+ views in one month (vendor-reported). (Syllaby case study)
|
 |
Guide 5 of 5
A kids boutique reported a 50% improvement in order processing time after moving every store into one order queue (vendor-reported). (Goflow case study, Dec 16, 2024)
|
|
 |
Worth a look
Top AI posts on X |
Like counts as of Oct 25, 2026 (from X). Dates are Pacific time.
K |
@Kimi_Moonshot · Oct 23 · 28.7K likes |
● Top post |
“Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.”
|
O |
@OpenAI · Oct 21 · 14.4K likes |
“OpenAI models helped researchers… discover critical security flaws in PHP, an open-source programming language used to run an estimated 70% of websites.”
|
O |
@OpenAI · Oct 23 · 13.3K likes |
“We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead.”
|
A |
@AnthropicAI · Oct 22 · 9.7K likes |
“New on our Frontier Red Team blog: using Claude to formally verify software at large scale.”
|
C |
@cursor_ai · Oct 23 · 6.4K likes |
“Cursor can now control your computer. Agents can drive desktop apps like the iOS Simulator to test the features they build.”
|
From @MishalBrands this week | |
@MishalBrands
Get daily AI tools on X
Short, practical posts on new AI tools, real pricing and step-by-step guides.
|
|
 | mishal brands |
Some links are affiliate links (disclosure).
|
|
|
