Issue #2

Patch PHP. Check LiteLLM. Then Fix Your TV.

Friday, Oct 23, 2026

AI found a decade-old hole in PHP, LiteLLM shipped a poisoned release, and people are using coding agents to strip the ads off their TVs.

Welcome back. This was a security week first and a launch week second, so the two urgent items come first and the fun stuff follows.

PHP2Shell: AI-assisted research found critical remote code execution bugs in PHP itself. Patches shipped Wednesday and attacks have already started.
LiteLLM 1.107.0 was malicious. If you installed it on Wednesday, treat that machine as compromised.
Off the radar: a 40B open coding model that only uses about 2.5B parameters per token, and a tiny model that pretends to be a terminal.
Steal this: people are using Claude Code and Codex to strip the ads and bloat off old Android TVs.

Big Drops

PHP2Shell: patch PHP now (Oct 21)

PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 fix two critical bugs (CVE-2026-109618 and CVE-2026-109619) that can allow unauthenticated remote code execution. PHP says every installation serving web requests is at risk. Wiz found the bugs with help from Gemini 4 Argon and GPT-5.6 Sol, and says every PHP version from the last decade, back to 5.4, is hit by at least one of them. Amazon spotted an exploit attempt on its honeypots dated Oct 17, four days before disclosure, and now says the bugs are "being broadly exploited in the wild."

Good for: anyone running WordPress, Drupal, Joomla or any other PHP app. Update today. Config tweaks only reduce exposure, and PHP is clear that updating is the only full fix. OpenAI published a free version checker as a Codex skill. Read PHP's advisory, Wiz's write-up and Amazon's exploitation report.

LiteLLM 1.107.0 shipped a credential stealer (Oct 21)

The popular LLM gateway's release pipeline was hijacked. litellm==1.107.0 sat on PyPI for about two hours on Wednesday, and the v1.107.0, latest, main-latest and main-stable Docker tags pointed to the bad image until 14:20 UTC. The malware runs every time Python starts, even if nothing imports LiteLLM. It grabs SSH keys, cloud credentials, browser passwords, coding-agent tokens and the provider keys stored in a LiteLLM proxy. LiteLLM says the attacker got in using GitHub CLI tokens found in the logs of an internal LiteLLM deployment, and that it has seen Claude Code embed those tokens in response bodies that ended up in logs.

Good for: anyone who runs LiteLLM unpinned or on latest. Pin 1.106.2, rebuild the environment and rotate every key that machine could reach. A related fix from GitHub, out the same day: gh auth login --expiring (GitHub CLI v2.103.0, public preview) issues tokens that stop working within hours if they leak. See LiteLLM's incident report, SafeDep's analysis and GitHub's changelog.

Mercury 3 from Inception (Oct 19)

A diffusion LLM: it generates text by refining many tokens in parallel instead of strictly one at a time. Inception claims 1,068 tokens per second on NVIDIA GPUs, a 1M-token context window and quality on par with GPT-6 Luna and Claude Haiku 5.5. It costs $0.20/$0.70 per million input/output tokens, and is half that during launch. It's OpenAI API compatible and live on Inception's API, OpenRouter, Baseten and Vercel.

Good for: browser agents, voice bots and subagents where speed matters more than raw intelligence. All benchmarks are Inception's own. Inception's announcement.

Voxtral 3 from Mistral (Oct 20)

Mistral's new speech models plus a voice agent SDK that handles listening, reasoning, tool calls and speaking in one real-time pipeline. Transcription covers 50+ languages with speaker labels and end-of-turn detection, the new TTS voices cover 35+ languages, and "Pro Voice" adapts a voice from 30 minutes of audio. The 4B Mini transcription model costs $0.0035 per minute. You can use the EU-hosted API or run it on your own servers.

Good for: phone and support agents, especially in Europe or anywhere audio has to stay on your own infrastructure. Comparisons with Whisper and others are Mistral's own. Mistral's post.

VS Code 1.143: Compare models (Oct 19)

The new Compare models preview in the Agents window sends one task to two or more models, lets you mix Copilot, Claude and Codex harnesses, and can add a Judge model and a Synthesize step to merge the best answers. Turn it on with sessions.chat.compareAgents.enabled. The release also makes the Copilot harness the default for new sessions.

Good for: settling "which model is better at our codebase" arguments with your own tasks instead of leaderboards. Skip the new on-device model unless you have an NVIDIA GPU with 96 GB or more of memory. Release notes.

Off the Radar

Recursive Flint (Oct 21) is an open-weight (Apache-2.0) coding model from Recursive. It has 39.8B total parameters, with about 2.5B active per token, and a 262K-token context window. Recursive reports 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1. Good for: self-hosters who want a fast coding model. It's an 80 GB download that needs one B300 or H200 GPU per replica for now. On a Mac, the indie engine Kapitsa already runs it. Scores are self-reported.
WorldTUI (Oct 21) is a 360M-parameter model trained to be a terminal. There's no real shell behind it: every keypress goes in and the model predicts the next screen. It can run vim and tmux, play snake and usually remember files you made a few commands earlier. Good for: a five-minute "huh" and a peek at where learned software is heading. The 208 MB demo runs in the browser and works best in desktop Chrome or Edge.
qwen3-4b-deslop (Show HN, Oct 19) is Qwen3-4B fine-tuned on a MacBook Air to drop the usual AI writing tics. On 40 held-out prompts, the author's "slop score" fell from 121 to 12.5 per 1,000 words, and em dashes went to zero. Good for: local drafting through Ollama (a 2.5 GB download). The author admits the voice is flat and that it still invents names and places.
earshot (Oct 15) gives coding agents ears. It describes a sound file in words, finds a specific moment in a long recording and writes a second-by-second listening report on a whole video. It runs on CPU as a CLI, a Python API and an MCP server, and it's MIT-licensed. Good for: anyone having an agent build demos, launch videos or game audio. The descriptions are useful evidence, not a replacement for listening yourself.

Vibe-Coding Corner

fx0 (Oct 19) is an open-source design agent from Vercel Labs. You describe a page and it writes the whole thing as one HTML file, keeping a version for every turn. It starts from ten ready-made design styles, or you can upload your own. Good for: landing pages and mockups, plus a readable example if you want to build your own agent app on Vercel's libfx. Model costs bill through your Vercel AI Gateway account, and each turn is capped at 64 model steps so a looping agent can't run up a bill. Try it.
Beam (#1 on Product Hunt, Oct 19) lets you keep agents and builds running on a Mac mini or Studio and pull their windows onto the Mac in front of you as real windows, each with a Dock icon. Traffic goes directly between your Macs, encrypted. Good for: anyone running agents on a headless Mac who is tired of Screen Sharing. It needs two Macs on macOS Sonoma or later, Apple silicon or Intel. The free tier gives you 20 minutes a day. The yearly plan is $84 ($7 a month), and lifetime is $199.

Use Case Worth Stealing: Fix a Bloated Android TV with a Coding Agent

On Oct 17, a developer posted before-and-after photos of his parents' TV, which Claude Code had cleaned up over Wi-Fi. He says the post reached 7.4M views. The repo that came out of it, CleanYourTV, has passed 22,000 stars, and similar projects like TV Detox appeared within days. Here's the recipe:

On the TV, turn on Developer options and USB debugging. Your laptop has to be on the same Wi-Fi.
Paste the repo's prompt into Claude Code, Codex or Gemini CLI and give it your TV model.
The agent connects over ADB (Android's built-in debug bridge), lists every app and asks before each change. It disables apps rather than deleting them, keeps an undo list and installs a small, ad-free home screen.
Turn debugging off when it's done.

Good for: reviving an old Android or Google TV without buying new hardware. You will need access to Claude Code, Codex or Gemini CLI. The honest caveat comes from TV Detox's own prompt: a few TVs got stuck at the logo or kept restarting. Keep the restore script handy, and don't do this to a TV you can't afford to fiddle with.

Gaming Pick: Co-op Anything

Last week we covered REA, the "reverse engineer anything" repo. Now Co-op Anything (created Oct 16, MIT) points that kind of agent at single-player games. You tell Claude Code, Codex, Cursor or another agent which game you want to play with friends, and it follows the repo's playbook. The agent looks for an existing co-op mod and fixes it, or reverse engineers the game and builds one from scratch. Then it tests the mod with bots inside the game and packages it for a one-click install, so friends can join over Steam. The README suggests pairing it with REA for the binary work. Deus Ex co-op is listed as released, with Fallout: New Vegas and Ocarina of Time in progress. The repo hosts no game files and stays away from anti-cheat and online services.

Good for: modders and anyone with a stack of single-player favorites. Results will vary a lot by game. This is modding, not a button press.

Browse every tool we've covered at toolsmithweekly.com.

No sponsors this week. Nothing in this issue is paid placement.

Researched and drafted with AI help, checked by a human.