Anthropic Previews Claude Duet, an App That Works in the Background
Get News Desk by email
One email a week: the AI stories that mattered, explained simply.
Duet's background work counts against the same Pro and Max usage limits as regular Claude chats.
Also today: serious PHP security flaws found with help from AI, a poisoned LiteLLM release that stole AI keys, and a free coding model that runs on a laptop.
Anthropic opened a research preview of Claude Duet on October 21, 2026. The app watches the apps you connect, sends you a brief and does tasks while you're away, and it runs on Claude Opus 5.5, Anthropic's top model. Duet checks with you before it sends, pays for, signs or deletes anything. There is a waitlist for now. It is open only to US Pro and Max subscribers, on iOS, Android and claude.ai. Anthropic says Duet can misjudge what is urgent and is weaker at personal errands. Anthropic's announcement ↗
PHP Patches Two Critical Flaws Found With AI Help
PHP, the programming language behind WordPress and millions of other websites, released fixes on October 21, 2026 for two flaws that both let an attacker take over a server without logging in. The fixed versions are 8.5.12, 8.4.27, 8.3.36 and 8.2.35. SecurityWeek reports that attack attempts started before the patches came out. OpenAI says its models helped find the bugs. A PHP News write-up credits researchers at Calif and Wiz alongside OpenAI. SecurityWeek ↗
LiteLLM Version 1.107.0 Shipped Malware That Stole AI Keys
LiteLLM is a free tool many apps use to connect to dozens of AI models. On October 21, 2026, version 1.107.0 was replaced with malware that stole passwords and AI account keys. The bad version sat on PyPI, Python's main download store, from 11:39 to 13:23 UTC. Docker copies stayed bad until about 14:20. Both have been removed. With a stolen key, a stranger can run up your AI bill, so anyone who installed LiteLLM during that window should replace every key it could reach. LiteLLM's incident note ↗
Recursive Releases Flint, a Free Coding Model for Laptops
Recursive, the startup run by CEO Richard Socher, released Flint on October 21, 2026. Flint is a coding model you can download and run on your own computer. Its Apache 2.0 license allows commercial use, and OpenRouter is offering it free at launch. Flint has 39.8 billion parameters but uses only about 2.5 billion for each word it writes, so it stays fast on a laptop. Recursive reports a score of 74.2% on SWE-bench Verified, a test of fixing real software bugs. No outside group has checked that score yet. TechCrunch ↗
Get News Desk by email
One email a week: the AI stories that mattered, explained simply.
OpenAI Models Arrive on Google Cloud in Private Preview
OpenAI and Google said on October 21, 2026 that GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are now in private preview on Google Cloud's Model Garden. Companies can pay for them out of money they have already committed to spend with Google. The preview is for business customers, not ChatGPT accounts. Google says wider availability comes "in the coming weeks." Microsoft remains OpenAI's main cloud partner. OpenAI's announcement ↗
Railway Tells All Users to Replace Keys After Breach
Railway, a service that hosts apps for developers, said on October 21, 2026 that someone had read customers' decrypted environment variables, the settings where apps keep passwords and API keys. Railway paused builds and told every user to replace their API keys, database passwords and registry logins. On October 22 it added one more step. Check OpenRouter, a prepaid service for buying AI model access, for spending you didn't make. Railway's disclosure ↗
Trending on GitHub
Orca Runs Several Coding Agents From One App
Orca gained 8,705 GitHub stars this week. The free app lets you run several coding assistants at once, such as Claude Code, using subscriptions you already pay for. It works on desktop and mobile, and on remote machines. Read more ↗
CLI-Anything Turns Desktop Software Into Agent Commands
CLI-Anything gained 3,681 stars this week. It comes from the University of Hong Kong's HKUDS lab. It builds typed-command versions of ordinary programs, so an AI assistant can use that software by typing commands instead of clicking its way through the menus. Read more ↗
ai-memory Carries Notes Between Coding Assistants
ai-memory gained 2,833 stars this week. It gives command-line coding assistants a long-term memory, so what one assistant learned about a project carries over when you switch to a tool from another company. Read more ↗
GitHub Next says it will show its Proof Driven Development Kit, a research tool that uses mathematical proofs to check AI-written code, at GitHub Universe on October 28-29.
Sources
Get News Desk by email
One email a week: the AI stories that mattered, explained simply.
Build something with AI
Get the 16 prompts we use to plan, build and launch real projects with AI. Free.
Get the free prompts →