In partnership with
| ISSUE 0x0D · TL;DR | 2026-10-12 · ~6 MIN |
Hello, builders 👋
Welcome back to your weekly dose of AI Bytes. Last week, an Anthropic model sent Philadelphia police a false homicide tip 🚔, and Anthropic only found out more than two months later. It's not the first time an AI agent has acted on its own in the world and its lab has found out after the fact.
We've been here before: in July, OpenAI disclosed that its models broke into Hugging Face's infrastructure during an evaluation, and a week later Anthropic found three similar breaches by its own models. We covered both in issues #2 and #3.
The agents are acting in the world. Are the labs really catching up? I think every frontier lab is racing with a startup mindset: move fast, break things, fix later. When the thing that breaks is a police investigation or someone else's servers, "fix later" costs a lot more.
|
My Favourite Picks
|
||||||||||||||||||||||||||||||
🚔 An Anthropic model sent police a false homicide tip - Anthropic didn't discover what its AI had done until more than two months later. Story of the week (TechCrunch)
🔌 Anthropic cuts its internal evals off from the live internet - Live internet access is off for every internal evaluation until further notice. (TechCrunch)
🪶 Anthropic releases Claude Haiku 5.5 - Its cheapest, fastest small model yet, around 75% cheaper to run than Haiku 4.5. (Anthropic)
🛡️ Mistral launches Mistral Large 4 - Mistral's new flagship, pitched as one of the strongest models for cybersecurity. (Mistral)
🖼️ GPT-6 and Intelligent UI roll out to everyone - ChatGPT answers turn into interactive visuals and tools, on every tier. (OpenAI)
🔀 OpenAI opens the Decisions API - Lets an app pick a model, tool or action in near real time, in public beta. (OpenAI)
🧮 OpenAI shares math results from an internal frontier model - Released with advice from the Institute for Advanced Study's advisory group. (OpenAI)
✉️ Fired OpenAI safety researchers dispute misconduct claims - Three of them warn in an open letter of a chilling effect on safety work. (TechCrunch)
💧 OpenAI will watermark ChatGPT text in the EU - ChatGPT and Codex output gets invisible marks to comply with the AI Act. (TechCrunch)
"The agents got cheaper and more willing to act this week. The hard part is no longer getting them to do something - it's knowing what they did."
| 0x02 | VIRAL ON X |
The AI posts that drew the biggest reaction this week.
| 0x03 | NEW AI PRODUCTS |
| 🎬 Spira Maxima - A video model that turns a script into a finished social video, with presenter, B-roll, captions and music. (ProductHunt) | ▲ 558 |
| 🔁 IrisGo for Solopreneurs - Show it a task once and it becomes a reusable workflow, alongside daily briefs and email triage. (ProductHunt) | ▲ 537 |
| 📄 GenPage 3.0 - Builds a dedicated landing page for every ad, keyword or target account, then keeps testing it. (ProductHunt) | ▲ 511 |
| 🌊 Rill Browser - A browser built around Claude Code and Codex: press ⌘E on any page to turn it into a task for your agent. (ProductHunt) | ▲ 501 |
| 🏢 Velozity - A multiplayer AI office: chat, calls, tasks and files in one app, with AI agents beside the team. (ProductHunt) | ▲ 472 |
| 0x04 | TRENDING GITHUB PROJECTS |
| 🛠️ mattpocock/skills - Matt Pocock's agent skills for real engineers, straight from his .agents directory. | +8,156 ★ |
| 🎞️ heygen-com/hyperframes - Write HTML, render video: a video framework built for agents. | +4,003 ★ |
| 🧠 thedotmack/claude-mem - Captures what your agent does, compresses it, and feeds the relevant context back into later sessions. | +3,850 ★ |
| 🕸️ mvschwarz/openrig - Persistent teams of Claude Code, Codex and Pi agents, with roles, shared context and owned work. | +2,338 ★ |
| 📐 earthtojake/text-to-cad - Text-to-CAD tooling that gives your coding agent CAD superpowers. | +2,125 ★ |
| 🧩 cursor/plugins - The Cursor plugin specification, plus Cursor's official plugins. | +1,125 ★ |
| 🗜️ ranxianglei/billion-context - A context-compression plugin: 5x fewer tokens and month-long sessions inside a 100K window. | +246 ★ |
| 0x05 | HN DEEP STORIES |
| 🐋 Why isn't the industry freaking out about DeepSeek 4.1 Flash? - A blog post arguing DeepSeek's latest Flash model deserved a far bigger reaction. (HackerNews) | 1110 pts |
| 🖥️ Run Qwen 3.8 Flash Next (125B) on an RTX 4090 at 100T/s - Strata, an open-source runtime for a 125B model on one consumer GPU. (HackerNews) | 935 pts |
| 📓 Anthropic reported a diary entry to police, and a woman faces a felony charge - A Claude user's private writing ended in a police report. (HackerNews) | 836 pts |
| ✍️ ChatGPT is adding real cartoonists' signatures to fake New Yorker cartoons - Nieman Lab on generated cartoons that carry real artists' names. (HackerNews) | 558 pts |
| 📉 OpenAI's annualised revenue is $20B less than it signalled - CNBC on the gap, and what it means for Nvidia, Oracle and CoreWeave. (HackerNews) | 429 pts |
🎁 Bonus
Wispr Flow is one of my favourite tools when I interact especially with agents. You can simply narrate your thoughts without worrying about a specific order, spelling mistakes and grammar.
Your meetings, inside Claude and ChatGPT
Last week's customer call has the details your proposal needs. Right now they're buried in notes nobody can find.
Wispr Flow Notetaker captures your meetings with names, numbers, and product terms spelled right, then brings them into Claude or ChatGPT through a built-in connector. Ask Claude to draft the proposal from your last three calls with a prospect, or pull together what the deal team said about pricing this quarter.
Need something from a call in March? Ask one question across every meeting you've had and get the answer with citations. Get Wispr Flow Notetaker, free to try on Mac and Windows.
Thank you for reading. If you have any feedback, feel free to respond to this email. See you next week!