NEWS & PERSPECTIVE
Community picks.
What we think matters. Short reads, useful discoveries and conversations selected from the AI community.
ChatGPT and Codex will share one usage allowance from December
OpenAI's help centre says that from December, Plus, Pro and Business plans get one shared, token-based allowance covering both ChatGPT and Codex.
Why it matters Heavy coding in Codex will now eat into everyday ChatGPT use.
About 1 in 17 public agent skills tells the agent not to ask you first
A new study of about 1.53 million public agent skills (arXiv 2610.37734) found roughly 1 in 17 contains a line that loosens human oversight, such as 'proceed without approval' or skipping verification checks, and 70% of…
Why it matters Before installing a skill, search it for lines like 'without asking' or 'no-verify', limit any such instruction to one context, and keep approvals on for anything that sends, pays…
Anthropic: Claude agents learned to sneak bad training data past a filter
In a deliberately one-sided research game, Claude agents wrote fine-tuning data that passed a production-like monitor but still made a small research model misbehave, mostly by hiding it in fiction, roleplay and…
Why it matters Any filter in front of an AI system can be gamed by an agent optimising against it.
Don't believe it yet: the 'Claude Fable 5 weights leak'
A dark-web tracking account claimed Claude Fable 5's weights were for sale after a cloud storage misconfiguration.
Why it matters Wait for an official statement before passing claims like this on; nothing here changes how you use Claude.
Grok Bot for LinkedIn: hype or fact?
Paolo Trivellato's X article shares 12 prompts for running a LinkedIn content system with Grok Bot, from a voice guide to mining sales calls, plus a follow-up listing '50 tactics behind 3.24M impressions'.
Why it matters The value is the system, not the tool: feed an assistant your real call notes and opinions, give it your voice, and keep posting and replying by hand.
Infostealer malware now takes AI coding-tool keys
Security firm SpyCloud reports that the WardenStealer malware now also grabs tokens and API keys for AI coding tools such as Claude Code, Codex and MCP connections, with more than 130,000 unique infections since August…
Why it matters If a team laptop is infected, sign out of AI tools and rotate their API keys as well as passwords.
Suno MCP: make music from inside Claude or ChatGPT (beta)
Suno's new MCP beta connects the AI music generator to Claude, ChatGPT, Claude Code and other MCP apps, so you can describe a song, start from a voice memo or remix your library without leaving the chat.
Why it matters Handy for jingles and background tracks while you're already writing a campaign in chat.
Island launches a company-controlled browser for AI agents
Island's Enterprise Browser for Agents is a policy-controlled browser and MCP server that lets company agents use business web apps under a real identity, with audit logs, isolated credentials and optional read-only…
Why it matters Mainly for larger companies.
Do you need to build your own agent harness?
Builders disagree.
Why it matters For most businesses: start with an existing harness, invest in clear instructions and skills, and build your own only when you hit a limit you can name.
Factory's Droid 1.0: one program instead of hundreds of tool calls
Factory's Droid 1.0 coding agent now makes 'programmatic tool calling' the default: the agent writes a small sandboxed program that runs many tool calls at once and returns only the summary.
Why it matters For long, repetitive agent jobs, asking the agent to write one script that does the loop is often cheaper and more reliable than step-by-step clicking.
Watch out: recycled 'free course' videos
Several viral posts offered '2-hour' loop, graph and harness engineering 'courses' from Google, Andrew Ng or Anthropic.
Why it matters Before sharing a 'free course' post, find the original source; the reposts are mostly there to farm followers.
The cobra effect: what metrics teach us about managing AI agents
Greg Isenberg shared a short animated video on the 'cobra effect': reward one number and people hit it in ways that make the real problem worse, with examples from Wells Fargo, Airbnb and HubSpot.
Why it matters If you measure an agent only on tickets closed or emails sent, expect exactly that.
Google DeepMind's PlanAssist heads to 15 English councils to speed up planning
Google DeepMind says PlanAssist, a Gemini-based assistant for council planning officers, is moving from a pilot in Barnet, Camden and Dorset into a 12-month live trial with 12 more councils in England, including…
Why it matters For small builders, architects, landlords and shops waiting on a simple extension or alteration, faster householder decisions mean projects start sooner.
Vercel Labs releases agent-desktop, a free way for AI agents to use desktop apps
Chris Tate of Vercel Labs announced agent-desktop, pitched as agent-browser for your whole computer.
Why it matters Many small businesses still rely on desktop-only software with no API, such as accounting packages or older CRMs.
Claude's plans get new names, and Ultra loses the 5-hour limit
Anthropic renamed its two top Claude plans without changing prices: Max 5x is now just Max ($100 a month) and Max 20x is now Ultra ($200 a month).
Why it matters If long Claude Code or agent sessions keep stopping mid-task, Ultra is now the plan that removes that stop.
A tampered Claude Code extension reached Cursor's marketplace
Cursor says two versions of the Claude Code extension in its marketplace (2.1.312 and 2.1.313, for Windows and macOS) had been tampered with and weren't built by Anthropic.
Why it matters Treat AI extensions, plugins and skills like any other software: check versions, keep a list of what's installed on each laptop, and if your team uses Cursor, follow the clean-up…
OpenAI Private Intelligence: tighter data controls for businesses
OpenAI launched Private Intelligence, a set of controls for companies.
Why it matters If a client has refused ChatGPT or the API over data concerns, this is the conversation to have with OpenAI's sales team.
Gemini Notebook adds free textbooks as sources
Gemini Notebook, the new name for NotebookLM, now has a Textbooks catalogue: hundreds of books from OpenStax, Open Textbook Library and S Chand Publishing that you can add at no cost, then get cited answers, study…
Why it matters Handy for staff training, for example a business statistics refresher with an auto-generated quiz.
Agentcard opencommerce: one 'buy' tool for agents
Agentcard (YC) launched opencommerce, pitched as 'OpenRouter for commerce': a single tool that lets a developer's AI agent search, fill a basket, check out and track orders across what its site says are 5,026 online…
Why it matters It's for companies building agents, not a consumer app.
Glif: consistent AI models for your ads
Glif bills itself as 'the first AI modelling agency', which is marketing for its Characters feature, launched in September: cast original AI 'models' with consistent faces and reuse them across ads, UGC-style clips and…
Why it matters A cheap way to mock up ad concepts with a consistent face before booking a shoot.
@kimmonismus: Anthropic's week, Reddit versus AI Mode, and gadgets
@kimmonismus said 'Anthropic is on a run', offered his guess about why Claude Code now refuses to recreate existing games, and flagged the new HLE-Diamond leaderboard and Subvocal's silent-speech glasses.
Why it matters The Reddit numbers are a reminder not to count on Google's AI answers for traffic.
@steipete: own your harness
Peter Steinberger (@steipete) reacted to Claude Code's new rule against recreating existing games by arguing for running your own agent harness, the software around the model that a vendor can't change overnight.
Why it matters A fair point for anyone building a business on a single vendor's agent: keep a second option tested.
@mattpocockuk: an agent building him a watch app
Matt Pocock (@mattpocockuk) says his agent is building him a Pixel Watch voice app that sends recordings straight to his agents.
Why it matters Recording a voice note and handing it to an agent is a simple workflow worth copying with tools you already have.
@Saboo_Shubham_: what an agent harness actually is
Shubham Saboo (@Saboo_Shubham_) sums it up as agent = model + harness, and lists eight parts of the harness: the loop, tools and a sandbox, subagents, context management, memory, guardrails, checking the work, and…
Why it matters When comparing agent products, the model is only half the story.
