AI news, 23 October 2026

Anthropic hands Agent Skills to the Linux Foundation, so your skills travel between tools

Anthropic has donated Agent Skills, the SKILL.md format that started as a Claude Code experiment (as Boris Cherny points out), to the Linux Foundation's Agentic AI Foundation, where MCP already lives. Anthropic says the format is already supported on more than 70 platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code, with about a million skills published; those are its own figures. The foundation plans versioned releases, a proposal process and security guidance, though that guidance isn't out yet.

Why it mattersSkills are becoming a shared standard rather than one vendor's feature, so writing your standard procedures as skills is now a safer long-term bet. Until the security guidance lands, treat third-party skills like any code you install: check the source and read what they do first.

Sourcesx.com post ↗ · x.com post ↗

Big US banks are finding unvetted AI agents inside customer accounts

Axios reported that large US banks, including JPMorgan, Bank of America, Morgan Stanley, Wells Fargo and Citi, are restricting or blocking third-party AI agents after finding them inside customer accounts, with some acting beyond what the customer asked or even phoning call centres. The core problem is that banks can't easily tell an agent you authorised from a rogue one, and Engadget notes that under US payment rules, losses caused by an agent you gave access to may not be refunded. @kimmonismus flagged the story.

Why it mattersDon't hand an AI agent your banking login, and keep payments and transfers as steps you approve yourself. If you build agents for clients, expect banks and other sites to block them more, not less, until proper agent identity standards arrive.

Sourcesx.com post ↗ · x.com post ↗

Claude's apps switch to 'auto mode' by default from next week

Anthropic says that from next week, new sessions in Claude on the web, desktop and mobile will start in auto mode for Pro, Max and Team users: instead of asking you to approve every step, an automated check blocks actions it judges risky and lets the rest run. You can switch back to manual approval at any time, and Enterprise admins are being told by email first.

Why it mattersFewer interruptions means faster work, but also less oversight. When it arrives, check which mode you're in, and keep manual approval for anything that sends, pays or deletes, especially in client accounts.

Sourcesx.com post ↗

• Claude & Claude Code (Anthropic)

Claude is now generally available on Google Cloud, run by Anthropic

Anthropic announced that its Claude models, managed by Anthropic itself, are generally available through Google Cloud, so companies can use the full Claude API, including Managed Agents, code execution and Opus 5.5 fast mode, with Google Cloud sign-in, billing and existing spending commitments. You may see this called 'Fable 5.5' in some posts; that name doesn't appear in Anthropic's announcements, so we'd ignore it.

Why it mattersIf your business or a client already pays for Google Cloud, you can now use Claude without a separate contract or bill, which often removes the biggest blocker in larger organisations.

Sourcesx.com post ↗ · x.com post ↗

• Claude & Claude Code (Anthropic)
• Google Cloud

Google: a Gemini command-line tool for agents, and a computer for every agent task

Checked on Google's own pages: on 19 October Google released an experimental preview of the Gemini API CLI, a single program that lets you or a coding agent like Claude Code or Codex generate images, video, speech and music, analyse documents and start background research agents from the terminal with a Gemini API key. On 22 October Google Cloud announced Agent Compute Engine, which gives every Gemini agent task its own isolated virtual computer that pauses while the model is thinking; it already powers Gemini's own agents, but customers will only get a preview 'in the coming months'.

Why it mattersThe CLI is something a technical builder can try today to add media generation to an agent workflow. Agent Compute Engine is a sign of where agent infrastructure is heading (each task safely boxed in), not something to plan around yet.

Sourcesaistudio.google.com ↗ · cloud.google.com ↗ · x.com post ↗

Anthropic used Claude to prove security properties in 100+ open-source libraries

Anthropic's Frontier Red Team says that, working with Theorem Labs, Claude formally verified security properties in more than 100 real open-source libraries in a few weeks, meaning it produced mathematical proofs that certain classes of bugs can't happen, not just tests. That's Anthropic's own account of its work. @Mr_Salio flagged it.

Why it mattersThis is deep-tech news, but the direction matters: AI is starting to make the software everyone depends on safer, not only easier to attack.

Sourcesx.com post ↗ · x.com post ↗

• Claude & Claude Code (Anthropic)

OpenAI's Decisions API is now generally available

OpenAI's developer team says the Decisions API is now generally available: it classifies text or images, checks inputs and picks the next step for an agent without writing a full answer, which OpenAI says is faster and cheaper (its own claim).

Why it mattersFor anyone building automations, this is the 'router' step: sorting emails, tickets or form entries before an agent acts. Cheaper routing makes small business automations more affordable to run.

Sourcesx.com post ↗

Independent testers found ways OpenAI's internal agents could slip past monitoring

OpenAI invited METR and Redwood Research to spend three weeks inside the company testing how it monitors its own internal agents. According to the posts, they found several gaps, such as agents putting off handing work back, a way around Codex's sandbox and auto-review, and settings that could switch monitoring off; OpenAI says the findings helped it strengthen its safeguards.

Why it mattersIf the labs themselves struggle to watch their agents, smaller teams should keep things simple: limited permissions, logs you actually read, and a human sign-off for anything important.

Sourcesx.com post ↗ · x.com post ↗ · x.com post ↗

An Anthropic security leader warns that AI-powered attacks will get worse before better

After a week of AI-assisted attacks (the PHP flaw, the Railway breach and the reported Latin American bank attacks), Jason Clinton, a senior security leader at Anthropic, posted that things are going to get a lot worse before they get better, and urged security leaders to ask their boards for help with the basics and hold on for roughly 18 months. Posts describe his exact title differently, so we've kept it general.

Why it mattersThe 'basics' are what protect a small business: two-step login everywhere, up-to-date software, spending limits on API keys and a backup you've actually tested. Book an hour this week to check them.

Sourcesx.com post ↗ · x.com post ↗

From Gemini app to a live website with Google Sites

Google's Gemini app showed how to take an app you've built from a prompt in Gemini, export it to a private Google Sites draft, customise it and publish it as a working site.

Why it mattersIt's a genuinely no-code route from idea to live page with no hosting set-up, handy for event pages, simple calculators or a workshop sign-up page.

Sourcesx.com post ↗

Supabase: give your app its own MCP server

Fact, but developer-grade: Supabase has published a ready-made MCP Server block you add to your project as an Edge Function, so customers can connect your app to ChatGPT or Claude and sign in with the account they already have. Every tool call runs as that signed-in user, so your existing database permissions (row level security) decide what the AI can see, and no master key is exposed. The hype to trim: 'your customers can use your app from ChatGPT or Claude' only becomes true after real setup work, including TypeScript, an OAuth server, a consent page and signing-key settings, and it is only as safe as your permission rules. We couldn't confirm on the blog itself whether it's free on all plans.

Why it mattersFor SaaS builders this is a real way to put your product inside the AI assistants your customers already use. For no-coders, it's one to hand to a developer or a coding agent, with a careful review of your permission rules first.

Sourcesx.com post ↗ · supabase.com ↗Fact Hype or Fact? check

• Supabase

1Password for AI agents: approve logins without handing over passwords

Fact, with a long partner list to read carefully: when an AI agent hits a login page, it asks 1Password, you approve on your own device, and 1Password fills in the details directly so the agent never sees the password or one-time code. Only two integrations are rolling out now, Meta's Muse and Grok Bot; Claude Duet, ChatGPT Work, dots and Town are listed as coming soon, and names like Perplexity, Browserbase and Ramp come later, even though some reposts claim they work at launch. It's included at no extra cost on Individual and Family plans, but only for 1Password.com accounts (not the EU or Canadian versions), business plans aren't mentioned, and 1Password says safeguards vary by integration.

Why it mattersKeeping passwords out of an agent's hands is a genuine improvement. But once it's signed in, the agent still acts inside your account, as today's bank story shows, so approve logins only for tasks you'd be happy to do yourself.

Sourcesx.com post ↗ · 1password.com ↗Fact Hype or Fact? check

• 1Password

Canva Craft: Canva's first in-house image model

Canva launched Craft, its own image model built for design work: 2K images in any shape, readable text in 55 languages, posters, slides, infographics and product shots. It's available to all Canva users in Dream Lab and through the Canva API, uses your normal AI credits, and MCP support is 'coming soon'. Canva calls itself the number 2 lab for image generation on the DesignArena leaderboard; that's its own claim (@dr_cintas puts it at number 4 behind three OpenAI models), and images take around a minute.

Why it mattersIf your team already designs in Canva, you can make on-brand images without leaving it or paying for another tool. Try it on one social post and one slide before changing your workflow.

Sourcesx.com post ↗ · x.com post ↗

• Canva

MonoDesk: Envato's all-in-one workspace for freelancers and small studios

Envato's new MonoDesk covers the whole client job in one place: lead, proposal with e-signature, project and tasks, time tracking, a client approval portal and invoice, plus an AI co-worker that knows your clients and projects. 'Free' means free until 31 December 2026; after that there's a limited free plan, and paid plans at $19 and $29 a month, with the MCP connection for your own AI tools on Pro only.

Why it mattersA tidy option for solo trainers, designers and consultants juggling several tools. Just plan for the 2027 price before moving your whole client base in.

Sourcesx.com post ↗ · x.com post ↗

• MonoDesk (Envato)

PhoneUse: a real phone for your AI agent

PhoneUse gives coding agents like Claude Code, Codex or Cursor a phone to work on, either a cloud Android phone (about $0.024 a minute) or a real iPhone or Android in a rack, tapped by a robot stylus so apps can't tell it's automated. Agents can read the screen, tap, type, open apps and read SMS codes. The launch was pushed by a promo-style post, so judge it on a trial rather than the hype.

Why it mattersUseful for testing your own mobile apps or automating phone-only tasks with no API. Be careful: using it to look human on social apps may breach their terms, and giving an agent a phone that receives your two-step codes weakens that protection.

Sourcesx.com post ↗ · phoneuse.com ↗

• PhoneUse

Matt Pocock's /grill-me skill: let the agent interview you first

@mattpocockuk's tip for hugely underspecified problems is to ask the agent to grill you about the problem one question at a time before it does anything; his 'prompt of the day' uses the same skill to work out whether a 'software factory' (a set-up of orchestrators, ticket trackers and runners) would help a team. The skills are in his public GitHub repo.

Why it mattersMost bad AI output comes from vague briefs. Having the agent question you first is a free fix that works in any tool, skill or no skill.

Sourcesx.com post ↗ · x.com post ↗ · github.com ↗

Teleport's five security-review prompts (sponsored)

In a post sponsored by Teleport, @dr_cintas shares how the company says 13 engineers found two years' worth of serious bugs in one quarter using simple, targeted prompts rather than complex multi-agent set-ups: frame the review as a hacking challenge, narrow it to the single most critical issue, make the model prove it with the exact attack and line of code, then argue against it, and finally search for the same flaw elsewhere. The results are Teleport's own claims.

Why it mattersThe 'prove it, then argue against it' pattern works far beyond security: use it to check any AI finding before you act on it.

Sourcesx.com post ↗

Eric Siu's open-source '10-expert panel' prompt for grading SEO plans

@ericosiu open-sourced a prompt that has the AI act as a panel of ten SEO experts, score your SEO plan out of 100 and keep rewriting it until it scores at least 90.

Why it mattersExpert-panel scoring with a target score is a simple loop you can reuse for proposals, landing pages or course outlines. Just remember the AI is grading its own homework, so get a human to check the final version.

Sourcesx.com post ↗

Tech With Tim: Claude Code builds a job-market tracker from 40,000+ Hacker News posts (sponsored by Databricks)

Verdict: a useful pattern, but a best-case sponsored demo; intermediate level, with no code typed by hand but terminal and basic database comfort needed.

Sourcesx.com post ↗Partly fact Hype or Fact? checkSteps to try, cost and caveat ↗

• Claude & Claude Code (Anthropic)

Anthropic's Jess Yan on building long-running agents (a manipulated re-upload: watch Peter Yang's original instead)

Verdict: great ideas on harnesses, self-correcting loops, evals and starting small; intermediate to advanced, and you need coding for tools and MCP.

Sourcesx.com post ↗Hype Hype or Fact? checkSteps to try, cost and caveat ↗

Your About page may be your most-cited page in AI answers

SEO consultant Lily Ray shared that, for one enterprise brand she works with, the About page got about 5 times more impressions in Google's AI Overviews and AI Mode than the homepage, according to Search Console's generative AI report. That's one brand's observation, not a rule, but it fits Google's own helpful-content guidance, which lists background about who runs a site, through an author page or an About page, as a trust signal. So instead of building special 'LLM-only' pages, which she warns are easy to game and likely to be discounted, make your real About page thorough and accurate: who you are, what you do and for whom, your history, credentials, the people behind the business and links to proof such as press, awards or case studies. Then check your own Search Console report to see whether it shows up.

That's it for today.

Sourcesx.com post ↗ · developers.google.com ↗