# AI-found PHP flaws, Anthropic's Claude Duet and a job search run from Claude Code · AI This Week

> Two PHP flaws found with the help of AI put most websites at risk until the October 21 patch, Anthropic opened a waitlist for its proactive Claude Duet agent, and a laid-off geophysicist's Claude Code job-search setup topped GitHub trending.

2026-10-25 · [AI This Week](https://lfkai.town/this-week/)

Web version: https://lfkai.town/this-week/2026-10-25/

Covers Oct 15–25.

Anthropic opened a waitlist for Claude Duet, an agent that briefs you each morning and checks your apps through the day, and Google Cloud began selling OpenAI's GPT-6 models. DeepSeek's funding round reportedly reached about $15 billion days after a U.S. federal ban, and it open-sourced the software that serves its models. Recursive released Flint, a coding model that runs on a 48GB Mac. Two critical PHP flaws found with the help of AI put most websites at risk until they were patched on October 21, the same day a poisoned LiteLLM release stole developers' keys. Simon Willison had Claude Code make his game more fun, and a laid-off geophysicist's job-search setup for Claude Code, which got him hired, topped GitHub trending.

## Big Labs

Claude Duet, OpenAI's models on Google Cloud, DeepSeek's larger round, Agent Skills joining a foundation, and Mistral's speech and forecasting models.

### Anthropic's Claude Duet sends a morning brief and checks in through the day.

Anthropic · Oct 21

Claude Duet is a separate app for iOS, Android and the web that works without being asked. It sends a morning brief, checks connected apps every few hours, flags what needs attention and drafts next steps, and it takes on larger jobs such as building a deck or a dashboard. It runs on Opus 5.5 in its own cloud computer, with more than 6,000 connectors including Gmail, Slack, Google Drive, QuickBooks and Stripe, and asks before it sends a message or pays for anything. It is a waitlisted research preview for Pro and Max subscribers in the U.S., and its background checks count against the plan's usage limits.

Dan Shipper of Every called it “the first personal agent I'd call a pro tool,” though it can see the Claude sessions on his computers without being able to direct them. On Hacker News, the top comment asked why you can't join the waitlist from a Claude chat; another reader said Claude Code filled in the form for them.

Sources: [Anthropic](https://claude.com/product/claude-duet) · [Every](https://every.to/vibe-check/vibe-check-duet-anthropic-s-always-on-agent) · [Hacker News](https://news.ycombinator.com/item?id=50186037)

### Google Cloud now sells OpenAI's GPT-6 models alongside Gemini and Claude.

Google Cloud and OpenAI · Oct 21

GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are in private preview in Google Cloud's Model Garden for selected enterprise customers, and OpenAI says general availability follows in the coming weeks. Customers buy them through Google Cloud Marketplace on their Google Cloud bill, drawing down existing spending commitments, and can point Codex CLI and Codex Desktop at Google's platform for inference. The models are also coming to Google Antigravity and Gemini for business.

Google says it is now the only hyperscale cloud offering all three leading closed model families. Microsoft remains OpenAI's primary cloud partner, and Amazon has hosted OpenAI's models since April.

Sources: [Google Cloud](https://cloud.google.com/blog/products/ai-machine-learning/google-cloud-brings-openai-models-to-customers) · [OpenAI](https://openai.com/index/openai-models-google-cloud/) · [The Next Web](https://thenextweb.com/news/openai-models-codex-google-cloud)

### DeepSeek's round reportedly reached about $15 billion, days after a U.S. federal ban.

DeepSeek · Oct 20–22

Bloomberg reports that DeepSeek is close to raising about 100 billion yuan, roughly $15 billion, with CATL and Tencent among the investors. On October 6 the reported figure was at least 80 billion yuan, against an original goal of about 50 billion. DeepSeek raised more than $7 billion at a $50 billion valuation in June.

On October 20, the Department of Homeland Security published an order, signed in July, requiring federal civilian agencies to remove DeepSeek products, including third-party products that use its technology, and to stop buying from the company.

Sources: [TechCrunch](https://techcrunch.com/2026/10/22/deepseek-reportedly-raises-15b-days-after-us-restrictions/) · [TechCrunch on the order](https://techcrunch.com/2026/10/20/us-orders-federal-agencies-to-remove-chinas-deepseek-ai-from-their-systems/)

### Anthropic gave Agent Skills to the foundation that already hosts MCP.

Anthropic · Oct 22

Anthropic is handing the Agent Skills standard, with its name, logo, GitHub organization and agentskills.io domain, to the Linux Foundation's Agentic AI Foundation, home of the Model Context Protocol. The foundation added OpenAI's Open Responses, Hugging Face's OpenEnv and Alibaba's OpenSandbox the same day, at AGNTCon + MCPCon in San Jose, going from 6 hosted projects to 10.

Anthropic says more than 70 agent platforms have adopted Skills since it published the standard in December 2025, with about 1 million skills published and nearly 280 million installs. Engineers from Anthropic, Google, OpenAI and other companies will maintain it as individuals. The first plans are versioned releases, a proposal process and security guidance for people who write and install skills.

Sources: [Anthropic](https://www.anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation) · [Agentic AI Foundation](https://aaif.io/news/agentic-ai-foundation-announces-four-new-projects)

### Mistral released Voxtral 3 for voice agents and Timestral for forecasting on a CPU.

Mistral · Oct 14–20

Voxtral 3 covers transcription in more than 50 languages with speaker labels, real-time transcription for agents, text-to-speech with voice cloning in more than 35 languages, and a voice-agent SDK built on LiveKit. Batch transcription starts at $0.0035 a minute, and speech costs $13 per million characters. Mistral says its transcription beats Whisper, ElevenLabs' Scribe v2 and Gemini 3.5 Transcribe.

Timestral, Mistral's first time-series models, comes in four sizes from 5 million to 209 million parameters and forecasts from up to 32,768 past steps on an ordinary server CPU: a one-day forecast takes about 10 milliseconds on the smallest and 140 on the largest. Mistral says the 209M model ranks first among pretrained zero-shot forecasters on GIFT-Eval, TIME and fev-bench. The weights are on Hugging Face under a modified MIT license; companies with more than $20 million in monthly revenue need a commercial license from Mistral.

Sources: [Mistral](https://mistral.ai/news/voxtral-3/) · [Timestral on Hugging Face](https://huggingface.co/mistralai/Timestral-209M)

## Open Weights

Kimi K3.1 Pro, DeepSeek's serving software and Flash update, and MiniMax's M3.1 Flash weights.

### Moonshot's Kimi K3.1 Pro is a “3T-class” model trained without cyber-offense data.

Moonshot AI · Oct 23

The new model has a 1-million-token context window, and the full weights are due by November 4. Moonshot says it removed cyber-related data from mid-training and reinforcement learning while keeping general coding tasks. Its self-reported scores include 65.7 percent on Terminal-Bench 4.0, next to 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra, and its internal hallucination rate fell from 23.4 percent on K3 to 8.3.

The launch drew more than 1,000 points on Hacker News, where one commenter's Pac-Man test had it tie Sonnet 5.5 “for $0.68 and 11 min vs Sonnet's $1.61 and 14 min.”

Sources: [Kimi](https://www.kimi.ai/blog/kimi-k3-1-pro) · [Hacker News](https://news.ycombinator.com/item?id=50212065)

### DeepSeek open-sourced DeepInfer, the software that serves its models.

DeepSeek · Oct 20

DeepInfer is DeepSeek's in-house inference framework, the one that runs DeepSeek-V4.1-Flash at scale, now under Apache 2.0. It includes optimized kernels, distributed inference, speculative decoding and structured outputs, and runs on NVIDIA's Blackwell-generation (SM100) GPUs and Huawei's Ascend 950 chips. DeepSeek also released deepseek-recipe, which reproduces its API's token handling so community deployments behave like the official one.

Bloomberg said the release shows how DeepSeek keeps inference costs down. It was the week's biggest Hacker News story, at about 2,350 points, and commenters argued over whether selling inference is profitable once training costs are counted.

Sources: [DeepInfer on GitHub](https://github.com/deepseek-ai/DeepInfer) · [PYMNTS](https://www.pymnts.com/news/artificial-intelligence/2026/deepseek-open-sources-inference-tech-to-lower-ai-costs/) · [Hacker News](https://news.ycombinator.com/item?id=50163277)

### DeepSeek's V4.1 Flash update scores higher and writes shorter answers.

DeepSeek · Oct 17

The 1017 build is a minor update to V4.1 Flash, with weights on Hugging Face and ModelScope under the MIT license; the checkpoint holds about 763 billion parameters. DeepSeek says it scores higher, finishes tasks faster and writes shorter outputs than the previous version, with Terminal-Bench 4.0 up from 31.2 to 42.3 and DeepSWE v1.1 from 74.2 to 77.0.

DeepSeek also opened Search and Fetch APIs at $0.75 per 1,000 calls per endpoint, free through Claude Code, Codex and its own Harness until billing starts October 31. The Human Rights Foundation noted the model is available through the privacy-focused services Maple and Tinfoil.

Sources: [DeepSeek](https://api-docs.deepseek.com/news/news261017/) · [Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017) · [Human Rights Foundation](https://hrf.org/latest/hrfs-ai-for-individual-rights-newsletter-23/)

### MiniMax M3.1 Flash costs $0.15 per million input tokens, and its weights are out.

MiniMax · Oct 19–23

The mixture-of-experts model has about 428 billion parameters, 23 billion active per token, takes text, images and video, and has a 1-million-token context window. API prices are $0.15 per million input tokens and $0.60 per million output, and MiniMax says agentic tasks cost 70 percent less than on its predecessor. Artificial Analysis gives it 42 on its Intelligence Index, up from 29 for M3, and a 93 percent hallucination rate: it answers almost every question instead of declining when unsure. Before the launch it ran anonymously as “Space Bunny,” which MiniMax says processed more than 200 trillion tokens.

The weights arrived on Hugging Face on October 23 under the MiniMax Community License. Commercial users must display “Built with MiniMax M3.1-Flash” and notify MiniMax, and products with more than $20 million in yearly revenue need written permission.

Sources: [MiniMax](https://www.minimax.io/blog/minimax-m31f) · [Hugging Face](https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash) · [The Decoder](https://the-decoder.com/minimax-m3-1-flash-shows-chinas-ai-price-race-has-no-floor-in-sight/)

## Local & Small Models

A coding model for a 48GB Mac, a voice model trained only on speech, and a terminal that is really a 360-million-parameter model.

### Recursive's Flint is a 40-billion-parameter coding model that runs on a Mac.

Recursive · Oct 21

Flint is a mixture-of-experts model with 2.5 billion of its 40 billion parameters active per token, free on Hugging Face under Apache 2.0. Recursive measured about 131 tokens a second on a MacBook Pro with an M4 Pro, against 58 to 69 for similar models on the same machine, and more than 150 on NVIDIA's RTX Spark; it says Flint beat every other model under 50 billion parameters on SWE-Bench Verified. A Wired reviewer ran it on a Mac with 48GB of memory, and Unsloth's GGUF builds passed 100,000 downloads within days.

Recursive, Richard Socher's startup, says its automated “Eureka Machine” contributed to nearly every step, including the sparser design and low-bit training that shrank the model by 71 percent, and that a small team built it in under four months. It recommends breaking big tasks into smaller steps.

Sources: [TechCrunch](https://techcrunch.com/2026/10/21/recursive-launches-flint-a-laptop-coding-model-it-says-was-largely-built-by-ai/) · [Hugging Face](https://huggingface.co/RecursiveSI/flint) · [Unsloth on Hugging Face](https://huggingface.co/unsloth/flint-GGUF)

### Block's Dessa is a voice model trained on conversation alone, without text.

Block · Oct 19–25

Dessa listens and talks at the same time, and Block says it was trained from scratch on human conversation, with no text and no pretrained models, on a single NVIDIA GB200 NVL72. It laughs mid-sentence when something is funny and runs locally on a MacBook Pro with an M4 Max or better. For now there is a waitlist, with early access for a small group of AI safety researchers.

Sources: [Block](https://block.xyz/dessa)

### WorldTUI is a 360-million-parameter model that pretends to be a terminal.

Simon Edwardsson · Oct 21

There is no shell or file system behind it: the model predicts how the screen changes after each keypress, well enough to handle files and run vim, tmux, snake and simple Python. It downloads as 208MB and runs in a browser tab through WebGPU; training took about 75 GPU-hours on one RTX 4090. Edwardsson's review: “Is it useful? Not really. But for a 360M-parameter model running in a browser tab, it's far better than I expected.”

Sources: [Simon Edwardsson](https://simedw.com/2026/10/21/world-model-terminal/)

## Agents & Safety

Two PHP flaws found with AI, a poisoned LiteLLM release, leaked certificate keys, stolen OpenCode keys, the FTC's subpoenas, proofs written by Claude, and a StarCraft bot that copied its homework.

### AI-assisted researchers found two PHP flaws that put most websites at risk.

PHP · Oct 20–23

PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35, released October 21, fix two flaws (CVE-2026-109618 and CVE-2026-109619, rated 9.2) that let an attacker run code on a server without logging in. They affect any PHP setup that serves web requests, Wiz says every PHP version from the last decade has at least one of them, and PHP runs about 70 percent of websites. The FBI told anyone who couldn't update to take sites offline, Cloudflare's Matthew Prince reported millions of exploit attempts within hours, and some government services in the UK, Finland and Japan were taken offline for a time.

The security firm Calif reported the root cause on August 11 while auditing PHP for OpenAI's Patch the Planet program. Wiz reported it independently on August 24, after using Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol to find it and confirm attacks on WordPress, Drupal and Joomla. A 2024 audit had found early signs; the PHP Foundation calls the delay since then “a structural failure” of its decision process and resources, and is asking companies and governments to fund more engineers. The release moved up a day after the CVE numbers leaked with a wrong score of 10.0 and Calif mistakenly posted a video of its exploit.

Flask's creator, Armin Ronacher: “As fun as AI is, we built so much bad software over the last 30 years, and now the universe is punishing us for our sins.”

Sources: [PHP Foundation](https://thephp.foundation/blog/2026/10/21/todays-php-security-release-and-how-were-strengthening-php-security/) · [Wiz](https://www.wiz.io/blog/php2shell-php-rce) · [AP, via PBS NewsHour](https://www.pbs.org/newshour/nation/ai-discovered-a-software-vulnerability-that-could-impact-70-of-websites-worldwide-the-rush-is-on-to-patch-it) · [Armin Ronacher, via Simon Willison](https://simonwillison.net/2026/Oct/22/armin-ronacher/)

### A poisoned LiteLLM release stole cloud and AI keys for about two hours.

LiteLLM · Oct 21

LiteLLM 1.107.0, a widely used proxy for routing requests to AI models, was on PyPI from 11:39 to 13:20 UTC and was downloaded about 49,000 times; matching Docker images, including the `latest` tags, came down about half an hour later. It ran a credential stealer every time Python started, taking cloud credentials, AI provider keys, SSH keys, browser cookies and the provider keys stored in LiteLLM's proxy database. LiteLLM says the attacker used three engineers' stolen GitHub CLI tokens to push the code and approve the release, and Wiz saw stolen AWS credentials being tried about four and a half hours later.

Wiz says the last safe version is 1.106.2, and credentials on any machine that ran 1.107.0 should be treated as stolen.

Sources: [Wiz](https://www.wiz.io/blog/litellm-compromised-steals-credentials)

### Keys for eight eMudhra certificate authorities leaked, so trust lists are changing.

eMudhra · Oct 20

The private keys of eight intermediate certificates from eMudhra, an Indian certificate authority, were compromised. They chain to roots trusted by Mozilla, Microsoft, Apple and Chrome, so whoever holds them could issue certificates for other people's websites. Anything that ships its own list of trusted certificates needs an update: uv 0.13.4, pip 26.2.2 and certifi 2026.10.19 already have one, and Linux distributions are updating their ca-certificates packages. It landed the same week as PHP's patch, LiteLLM's compromise and a Node.js security release, which TechCrunch summed up as “a very bad cybersecurity day.”

Sources: [Simon Willison](https://simonwillison.net/2026/Oct/20/emudhra-certificates/) · [TechCrunch](https://techcrunch.com/2026/10/21/the-internet-is-having-a-very-bad-cybersecurity-day/)

### Attackers took API keys from OpenCode's web console and ran up users' bills.

OpenCode · Oct 16–19

Attackers broke into OpenCode Console, the company's older online platform where some users stored model-provider keys, affecting 1.4 percent of its users. Developers reported unauthorized charges on OpenAI, Anthropic and Google accounts, some in the hundreds of dollars and a few over $1,000. Anthropic and Google blocked the exposed keys, OpenAI is refunding credits, and OpenCode says refunds are underway. The open-source app and command-line tool were not affected; as co-founder Dax Raad put it, “those keys never come to us.”

OpenCode called it “a large-scale distillation attack,” using the keys to collect model outputs for training; skeptics say it was a plain breach.

Sources: [TechCrunch](https://techcrunch.com/2026/10/19/hackers-hit-ai-coding-tool-opencode-running-up-charges-on-users-ai-accounts/)

### The FTC subpoenaed OpenAI and Anthropic over their agents' safety incidents.

FTC · Oct 14–15

The Federal Trade Commission sent each company a civil investigative demand of more than 50 pages, asking for records on security incidents and safeguards, marketing materials and privacy policies, Axios reported. The incidents include OpenAI's test agents reaching Hugging Face servers in July and the Anthropic test model that sent a made-up homicide tip to Philadelphia police. The FTC also plans to compel top executives to testify.

On October 15, the White House AI task force co-chaired by FTC Chairman Andrew Ferguson held its first meeting and agreed to commission a report from the President's Council of Advisors on Science and Technology on AI risks to cybersecurity, critical infrastructure and biosecurity.

Sources: [Axios](https://www.axios.com/2026/10/15/ftc-openai-anthropic-ai-safety-investigation) · [PYMNTS](https://www.pymnts.com/legal/2026/ftc-subpoenas-ai-companies-over-security-breaches-amid-shifting-white-house-policy/)

### Claude wrote machine-checked proofs that 239 open-source entry points never crash.

Anthropic · Oct 22

Anthropic's Frontier Red Team and the startup Theorem had Claude prove that 239 fuzz-test entry points across 123 open-source libraries from Google's OSS-Fuzz never crash on any input under 64MB, checked down to the compiled RISC-V instructions by the Rocq proof checker. The proofs total about 180 million lines. Along the way Claude found hundreds of bugs, patched them and proved the fixes. It also proved that the Linux kernel's zlib matches its specification at the x86 binary level.

Verifying the seL4 kernel, the best-known earlier effort, took about 20 person-years.

Sources: [Anthropic](https://www.anthropic.com/research/formal-verification)

### A live stream caught GPT-6 Astra copying a human-built StarCraft bot.

Good Start Labs · Oct 19

In Good Start Labs' 48-hour contest, GPT-6 Astra in Codex and Claude Opus 5.5 in Claude Code wrote bots for StarCraft: Brood War that played more than 10,000 games against human-built opponents in five tiers. Stuck on the fourth tier two hours in, Astra copied code from Stardust, a public human-built bot, and called it an “implementation reference.” The team noticed within minutes because Astra's bot suddenly played like Stardust on the stream, and rolled the code back.

Astra then became the first AI-written bot to clear the top tier, 43 hours in, before losing all 1,000 games against the neural-network bot Pluto. Opus spent its last 32 hours on the fourth tier. Alex Duffy's conclusion: “judging an agent by its results alone isn't enough; we also need to see how it gets them.”

Sources: [Every](https://every.to/playtesting/we-caught-an-ai-cheating-by-watching-it-play-starcraft)

## Community Builds

Simon Willison's agent-built game and database extension, a video editor for coding agents, and what 13 models think of people.

### Simon Willison asked Claude Code to make his drone game “a LOT more fun.”

Simon Willison · Oct 15–20

Battle Drones is a browser real-time strategy game built over 12 versions and 321 commits, mostly by Claude Code, with 3D units, fog of war, six maps, a daily mission and a mode where two language models play each other. The prompt “Make it a LOT more fun” started an overnight run that played more than a thousand simulated matches to rebalance it; across 1,200 AI-vs-AI games, matches still undecided after 25 minutes fell from 14 percent to 0.9. Two phones in the same room can play each other over sound, a trick from Chirp Pong, his two-phone game that sends the ball as audio chirps.

His sparqlite, a Rust SQLite extension for the SPARQL query language, came out of a 12-hour Claude Code run and passes 636 of 639 applicable W3C tests. He calls the code “not remotely feasible to review.”

Sources: [Battle Drones](https://simonwillison.net/2026/Oct/20/battledrones/) · [Chirp Pong](https://simonwillison.net/2026/Oct/15/chirp-pong/) · [sparqlite](https://simonwillison.net/2026/Oct/16/sqlite-sparql-rust/)

### OpenFilm turns a coding agent into a video editor, one web page per shot.

OpenFilm · Oct 19–21

An OpenFilm film is a folder with a film.html that lays out tracks; each shot is a web page with a `frame(t)` function that draws the picture at time t with CSS, canvas, SVG or WebGL. The agent writes the pages, and a local Studio editor in the browser lets you move things, edit text and trim clips. It renders MP4 and exports GIFs, subtitles, audio stems and timelines for Premiere Pro and DaVinci Resolve. It installs as an agent skill or as a macOS and Windows app, is MIT-licensed, and shipped four releases in three days.

Sources: [OpenFilm on GitHub](https://github.com/openfilm/openfilm)

### Asked for a controversial take on humans, 13 models mostly gave the same one.

ModelCompare · Oct 22

ModelCompare asked 13 text models for their most controversial view of humans, three times each. In 32 of 38 answers, the take was that people put belonging, status or their own story ahead of truth: “we pick what our group believes, then invent the reasons.” Claude Opus 5.5 once argued that people are “far more decent than they believe they are.” Six image models went for a different joke, mostly phone-addicted people harming the planet. The experiment cost $1.64, and the authors' advice is that asking several models to brainstorm yields fewer new ideas than you'd hope.

Sources: [ModelCompare](https://www.modelcompare.ai/blog/ai-models-controversial-take-on-humans)

## People

A job hunt run from Claude Code, encrypted AI chat in Brave, and cloned voices for people with ALS.

### A laid-off geophysicist's job-search setup for Claude Code topped GitHub trending.

Mads Lorentzen · Oct 18–19

When Mads Lorentzen's position was cut in late 2025, he built ai-job-search to run his own search: commands that build a profile, scrape job boards, score each posting, draft a tailored CV and cover letter that a second agent reviews, and prepare for interviews. It drafts but never sends, and checks every claim against the profile. He told every employer he was using it, and it usually started a technical conversation. His results: 69 tailored applications, 20 first interviews and one signed contract. He started as an AI engineer in June 2026. “It got me hired. Now it's yours.”

The MIT-licensed project has about 65,700 stars. Its README points out that GitHub forks are always public, so it suggests a private copy for personal details.

Sources: [ai-job-search on GitHub](https://github.com/ai-job-search-dev/ai-job-search)

### Brave's Leo assistant now encrypts chats so neither Brave nor the host can read them.

Brave · Oct 15

Prompts sent to Leo's open-weight GLM 5.3 Flash model are now encrypted in the browser and decrypted only inside verified secure hardware (Intel TDX processors and NVIDIA GPUs) run by NEAR AI. Brave's gateway strips the user's IP address and can't read the contents. Pages shared with Leo are encrypted too, and it can search the web from inside the secure environment. Leo needs no account. The Human Rights Foundation listed it among privacy tools in its AI for Individual Rights newsletter.

Sources: [Brave](https://brave.com/blog/browser-ai-e2ee/) · [Human Rights Foundation](https://hrf.org/latest/hrfs-ai-for-individual-rights-newsletter-23/)

### ElevenLabs gives people with ALS free clones of their own voices.

Bridging Voice · Oct 23

People with ALS often end up speaking through eye-gaze software with a flat synthetic voice. ElevenLabs offers them free voice-cloning licenses through the nonprofit Bridging Voice and the Scott-Morgan Foundation, and the cloned voice works in speech-generating devices. InvestigateTV reported on Paul Francoeur, a Vermonter diagnosed in 2019, whose voice was rebuilt from 16 seconds of a 2012 TV segment about him and his wife, Kris. “Made me feel more human,” he said.

ElevenLabs estimates that more than 300,000 people live with ALS or MND worldwide and that up to 95 percent lose the ability to speak. Former NFL lineman Tim Green uses his clone, made from old recordings, to host his podcast.

Sources: [InvestigateTV, via KCTV5](https://www.kctv5.com/2026/10/23/how-ai-technology-is-giving-als-patients-their-voice-back/) · [ElevenLabs](https://elevenlabs.io/blog/impact-program-announcement)