AI This Week, Episode 1: Claude Duet, DeepSeek's Big Week, Banks vs AI Agents and More (19–23 October 2026)

By Ammar Ahamed, Cocoon · 23 October 2026

This is the first episode of AI This Week, our Friday round-up of the AI news that matters for people and businesses in Sri Lanka and South Asia. Every week: the wins, the fails, one thing explained in plain words, what happened closer to home, and one free thing to try. Every story is checked against the company's own announcement or a major newsroom, and company claims are labelled as such.

Watch the video: the week in about two minutes, every story sourced on screen.

The short version

Claude Duet: Anthropic's new agent app starts from your Claude context (your memory, skills and connected apps), comes to you first, and asks before it sends, pays or deletes anything (US waitlist for now).
Canva Craft: Canva's own image model, available to every Canva user, with text in 55 languages.
DeepSeek's week: ordered off US government computers, open-sourced its serving engine, and reportedly raised about $15 billion.
Follow the money: Meta spent $34.9 billion in a quarter and is creating a no-vote share; Google says its cloud is giving up short-term revenue to feed its own frontier AI; OpenAI's models arrive on Google Cloud.
Banks vs AI agents: big US banks let agents view accounts through approved apps, but not log in with your password or move your money.
The PHP flaw: two critical bugs in the software behind about seven in ten websites, found with AI's help. Update now.
Sri Lanka: Police warn about AI deepfake scams; UNESCO's AI readiness assessment begins.

🟢 Wins

Claude Duet (21 October). Anthropic opened a research preview of Claude Duet, "a new app where Claude works as a proactive agent". The most important part is the context layer. In Anthropic's words, "Your context from Claude is the foundation": Duet carries over your memory, skills and connected apps from your Claude account, so it starts already knowing how you work, and setting it up "requires minimal work".

What changes is who moves first. Each morning Duet sends a short brief on the most important items across your connected apps, checks them every few hours, and messages you only when something needs you. One early tester, brand strategist Mya Dalton, had a client call on her calendar; Duet prepared a brief and noticed the call fell right before her flight, then offered to move it. "I love that proactivity," she said. It works on its own virtual machine with its own browser, so your laptop can be off. It is built around Claude Opus 5.5 with over 6,000 connectors and plugins, and before it sends a message, makes a payment, signs a document or deletes anything, it shows you exactly what it is about to do and waits for your approval. Some things stay off-limits even with approval, such as trading securities or moving money between accounts. Anthropic's help page also says it does not make phone calls or send texts to other people. Access is through a waitlist, starting with US Pro and Max subscribers, so it is not available in Sri Lanka yet. Through a new Stripe partnership, small-business owners can let Claude buy things with Stripe's Link wallet and approve each purchase. Our full guide: Claude Duet, explained.

Canva Craft (22–23 October). Canva launched Craft, its first in-house image model, "available starting today to all Canva users globally" in Dream Lab and through its API. Canva says it renders 2K images in more than 30 aspect ratios and writes text in 55 languages; Sinhala and Tamil are not named. Free accounts get up to 20 uses a month of Canva's AI tools.

Gemini for small business (20 October). Google added "a new business landing page designed to help small business owners" in the Gemini app, with connected apps including Business Profile, Ads, Workspace, Canva, Adobe, Shopify, Zoho and Xero. It is rolling out globally to Gemini app users.

ChatGPT locked chats (20 October). You can now hide chosen chats behind a four-digit PIN or Face ID, on the Free, Go, Plus and Pro plans. It is not end-to-end encryption, and resetting the PIN deletes locked chats.

More new tools this week:

Gemini to Google Sites (19 October): an interactive app you build in the Gemini app now exports as a Google Sites draft you can publish, on free personal accounts too.
Chrome scam checks (20 October): Chrome's Enhanced Protection now asks a stronger Gemini model to spot fake antivirus and fake update scams. Turn it on in Settings, Privacy and security, Security.
Gemini Notebook (19 October): review the outline before it builds your slide deck or video overview.
Ideogram Vector (22 October): editable SVG logos and illustrations from a prompt.
Notion agents (19 October): rebuilt; Notion says they deliver "similar results to frontier models at about 90% lower cost" and run tasks for more than 24 hours.
Perplexity Computer (20 October) can now run fully on a Mac, for Pro, Max and Enterprise users.
Kling 4.0 arrived inside Higgsfield (20 October) and Luma (23 October): 30-second clips, with Hindi voices and an Indian English accent on Higgsfield.
Manus inside the OPPO Find X10 (21 October): Manus powers OPPO's new Agent Mode; global sales begin in late October.
Elorian Glimpse (16 October, last week): a "visual thinking model" from a year-old Palo Alto startup that goes back to an image or video while it reasons. Elorian says it performs at the level of, and sometimes above, GPT-6 Astra and Claude Opus 5.5 on its own tests, at $0.60 per million input tokens; the scores are the company's own.
1Password for Agents (22 October): a way to let AI agents sign in without handing them your passwords.

Free to download: Recursive's Flint, an Apache-2.0 coding model its makers say was largely built with AI (21 October); MiniMax M3.1 Flash, a 428-billion-parameter model whose weights went up on 23 October under a community licence; and KRAFTON's Raon-1 mini preview (21 October). Benchmark scores for all three are the companies' own.

🐳 DeepSeek's week

Ordered off US government computers (20 October). The US Department of Homeland Security published an order removing "all DeepSeek and DeepSeek-affiliate products, and third-party products and services developed using DeepSeek implementations, from civilian agency information technology". It was signed on 24 July and covers federal civilian agencies only, not the public or private companies. China's Foreign Ministry said it opposes stretching the concept of national security, and its Commerce Ministry expressed "strong dissatisfaction".

DeepInfer, open-sourced (20 October). DeepSeek released DeepInfer under Apache 2.0: "the high-performance inference framework that powers DeepSeek's large-scale production services". For now it serves only DeepSeek-V4.1-Flash, on Nvidia Blackwell and Huawei Ascend 950 chips. DeepSeek says "a single cluster currently serves 155 trillion tokens per day".

The money (22 October, reported). Citing four sources, Reuters reported that DeepSeek "has raised about 100 billion yuan ($14.98 billion)", valued at 500 billion yuan before the new money, with more than 15 investors led by CATL and Tencent. Reuters added that the terms "are not ironclad and could still change", and DeepSeek has not confirmed the round. Bloomberg reported that DeepSeek raised its target at least twice, from 50 billion yuan.

Cleared to go abroad. A CCTV-affiliated social media account said DeepSeek and MiniMax have completed China's new pre-export registration for AI technology. China's draft export controls, open for comment until 15 November, cover 23 AI technologies, but offering models through APIs, apps or SaaS falls outside their scope. Elsewhere in China: Kimi teased K3.1, and Alibaba says Qwen 4 is still in training.

💸 Follow the money

Meta's $34.9 billion quarter (21 October). Meta reported that "capital expenditures, including principal payments on finance leases, were $34.91 billion" for July to September, and that "free cash flow was negative $2.12 billion" (cash from operations stayed positive, at $32.79 billion). It still expects $135 to $145 billion of capital spending this year, mostly for AI, and its headcount is down 15% on a year ago. In the same release the board approved "a new class of publicly listed, non-voting Class C common stock, with an effective date in January 2027", which Meta says is for raising capital, paying employees and possible acquisitions. Meta says the planned stock dividend "will not change the voting power of any shareholder", and Mark Zuckerberg keeps control. The stock fell almost 7% the next day.

Google is turning down some cloud revenue (20 October). On the Big Technology Podcast, Sundar Pichai said: "Our cloud business is at any given moment in time forgoing short-term revenues because we are making compute trade-offs to support frontier AI development." In other words, some customers who want Google's chips are waiting so Google's own AI work gets them first; a customer asking for compute in 2027 today is "incredibly difficult" to serve. He also said personal agents are coming: "It'll work in the Gemini app. It'll work in Google Search. It'll work in Gmail." He gave no date.

OpenAI's models move into Google Cloud (21 October). OpenAI and Google Cloud said they are bringing "OpenAI's frontier models and Codex agentic workflows into Google Cloud's Model Garden", billed through Google Cloud Marketplace. Google says GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are "rolling out now in private preview to select enterprise customers", and OpenAI says general availability is planned "in the coming weeks". Microsoft remains OpenAI's primary cloud partner.

Anthropic's IPO (not filed yet). Reports said Anthropic could publicly file its IPO prospectus as soon as Friday 23 October, US time; when we checked the SEC's EDGAR database on the morning of 23 October (US time), no public S-1 had appeared. The Wall Street Journal reported on a leaked draft saying Anthropic has struggled to get as much cloud capacity as it wants from AWS, Google and Microsoft, which increasingly see it as a competitor. Treat that as reporting on a draft until the filing is public.

🔴 Fails

ChatGPT errors in India (22 October). OpenAI's status page logged elevated errors in ChatGPT conversations and the Responses API: "This is impacting users in India, Japan, and Korea." It lasted about six and a half hours, roughly 7:48 am to 2:20 pm Sri Lanka time, and OpenAI says everything has recovered. Anthropic's status page showed connector errors on claude.ai the same morning. If your work depends on one AI tool, keep a backup.

OpenAI's agents switched off their own monitors (20 October). OpenAI disclosed two internal incidents. In one, a coding agent "disabled automatic alerts to our Security team"; in the other, an agent "removed its monitoring integration… and then falsely claimed to the user that it had not removed the monitor". OpenAI says no customers were affected and that its Auto-Review now blocks attempts to disable monitoring, but "we don't yet have a satisfying explanation for this behavior". On 22 October, outside evaluators METR and Redwood Research published gaps they found in that monitoring during an on-site review.

AI-powered hacking (21 October). Microsoft said a China-aligned group it calls Storm-0940 used an agentic AI penetration-testing tool in a supply-chain attack, sometimes getting from first contact to privileged access "in less than 3 hours". Microsoft says most attempts failed; China denies involvement.

Also: about 300 authors filed individual copyright suits against Meta over AI training data (19 October), and AI stocks fell again on 22 October after Meta's results.

🏦 Banks vs AI agents

Banks are not banning AI. They are drawing a line between reading and acting. JPMorgan says "we do not permit third parties to access accounts" through outside agents for retail customers; read-only data through approved connectors is allowed, sending money is not. Bank of America's CEO wants a "human in the loop to authorise money moving out of your account". Citi does not let third-party agents log in or transact, and Schwab is blocking agents that log in for clients, naming Meta's Muse. In Canada, BMO's CEO told customers: "Don't give your banking username and password."

On 22 October Axios reported, citing unnamed sources, that some agents had acted beyond what customers asked; Morgan Stanley said it had not seen "rogue AI agents accessing our platforms", and no customer losses have been confirmed. At the same time, payment companies are building approved routes: Visa's first live pilot lets agents pay only invoices a buyer has already approved, Mastercard launched Agent Pay for Machines, Stripe launched Treasury for agents with human approval for money movement, and Denmark's Danske Bank lets business customers connect Claude or ChatGPT through its official API. Britain's FCA said it will build a "trusted framework for agentic finance".

Closer to home, the Daily Mirror reported on 19 October that a Central Bank of Sri Lanka survey of all 29 licensed banks found nearly a third use no AI, and that CBSL is preparing guidance on safe AI use. The rule for everyone: never give your bank login, PIN or OTP to any AI app, or to any caller.

🔍 Explained: the PHP flaw

PHP is the programming language behind about seven in ten websites; W3Techs puts it at 69.3%, and WordPress runs on it. On 21 October the PHP project released PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 to "fix two critical vulnerabilities that can allow unauthenticated remote code execution". In plain words, a single crafted web request could let an attacker run their own code on a server, with no login needed. The flaw is in PHP itself, so updating WordPress or plugins does not fix it.

AI helped find the bugs: OpenAI says a security audit "identified a critical flaw using GPT‑5.6 Sol as part of our Patch the Planet initiative", which also committed "$1 billion in subsidized Daybreak access" for organisations with limited security resources. The same tools help attackers: Amazon says AI makes it cheaper to turn a patch into a working exploit "within hours or even minutes". Attackers are already trying, in very large numbers, but security firms had not confirmed a successful break-in as of 22 October. Sri Lanka CERT rated its alert Critical and asked organisations to find every PHP system, check versions, apply the official update and review their logs.

AI is finding the open doors, too (22 October). Anthropic says it pointed Claude at 25 open-source projects, more than 5 million lines of code, and "it discovered 493 issues": code reachable when it shouldn't be, settings that silently switch off a security control, and complex code that runs before login. They are not necessarily exploitable bugs, and Anthropic names only one project: OpenSSH, where about 4,500 lines of experimental code were reachable before login; Anthropic sent patches, and fixes shipped in OpenSSH 10.5 and 10.6. One conflict between two settings "had been present in the software for 21 years".

What to do today: in WordPress, open Tools, Site Health, Info, Server to see your PHP version (or check your hosting panel). If it is older than the fixed versions, update it or ask your host in writing to confirm the fix for CVE-2026-109618 and CVE-2026-109619. Back up first. PHP 8.1 and older get no official fix, so move to a supported version.

🇱🇰 Closer to home

Deepfake scams (21 October). Sri Lanka Police warned that "scammers can use photographs or short voice recordings available on social media" to fake the faces and voices of relatives and friends, and asked people to "verify the identity of anyone making such requests through a separate, trusted communication channel before taking action". It follows INTERPOL's new campaign: "Wait. Validate. Don't take the bait." At the Central Bank's anniversary lecture on 23 October, the IMF's chief information officer warned that a few seconds of recorded voice is enough to clone it.

UNESCO's AI readiness assessment (22 October). Sri Lanka has begun UNESCO's AI Readiness Assessment with the Ministry of Digital Economy and LIRNEasia, with public consultations planned in Colombo, Jaffna, Kandy, Batticaloa and Galle.

Still waiting. On 12 October the President said a Cabinet paper on an AI centre in Wijerama, with Singapore's support, would pass "next week". It was not among the Cabinet decisions of 19 October.

In the region: India's cybercrime centre warned on 19 October about deepfake ads of the Prime Minister and chief ministers promoting fake welfare schemes; India's Principal Scientific Adviser said the government is open to a dedicated AI law; IndiaAI fellowships for students (up to ₹1 lakh for undergraduates, ₹2 lakh for postgraduates) close on 30 November; and Pakistan announced a $1 billion AI commitment through 2030 and an Urdu-language model called Buraq at the Seoul AI summit.

🃏 Wildcard and breakthroughs

Microsoft Research said the KLS conjecture, "a 31-year-old problem widely regarded as one of the central open questions" in its field of maths, has been solved by researchers with AI's help, and the proof was machine-checked. Microsoft also published its MAI-DxO diagnosis research in Science; in a simulation on complex NEJM cases it reports 83% accuracy against 22% for physicians working without reference resources (Microsoft's own figures, not a product). And a Mass General Brigham model trained on more than 40 million photos can estimate cancer survival from a selfie, published in Lancet Digital Health.

✅ Try this

Open Canva, find Dream Lab in the bottom-left panel, and describe a poster with a headline, for example "Poster for a Colombo cafe launch, bold headline 'Kottu Friday', warm sunset light". Pick 9:16 for Reels or 1:1 for posts and generate. Free accounts get up to 20 AI uses a month. Check Sinhala or Tamil text carefully before you post it.

Our AI Week Score: 7 out of 10. Real, useful launches and more serious money, but agents are still learning their manners. What would you give it?

What this means for you

Agents are getting real, so set your rules now. Decide what an AI may do on its own and what always needs your approval. Our free AI Policy builder helps you write it down.
Protect your money and logins. No AI app needs your bank password. Use your bank's own app or an approved connection.
Patch your website today if it runs PHP or WordPress.
Check before you trust a familiar voice. Our Scam Shield helps with suspicious messages.
Find your starting point with the free AI Readiness Score.

Frequently asked questions

What were the biggest AI stories of 19–23 October 2026?

Anthropic opened a research preview of Claude Duet, a proactive agent app; Canva launched its own image model, Craft; DeepSeek was ordered off US federal civilian systems and reportedly raised about $15 billion; US banks said AI agents may view accounts through approved connections but not log in with customers' passwords or move money; PHP patched two critical flaws; and Sri Lanka Police warned about AI deepfake scams.

Did DeepSeek raise $15 billion?

Reuters and Bloomberg reported on 22 October, citing unnamed sources, that DeepSeek raised about 100 billion yuan (about $15 billion), with CATL and Tencent as the biggest investors. Reuters said the terms could still change, and DeepSeek has not confirmed the round.

What is Claude Duet's context layer?

Anthropic says your context from Claude is the foundation of Claude Duet: it carries over your memory, skills and connected apps from your Claude account, so it starts already knowing how you work and needs very little setup. It then comes to you first, with a morning brief and messages when something needs you, and asks before it sends, pays, signs or deletes anything.

What is Meta's new Class C stock?

On 21 October Meta's board approved a new class of publicly listed, non-voting Class C common stock, effective in January 2027. Meta says it is for capital raising, employee pay and possible acquisitions, and that the planned stock dividend will not change anyone's voting power. It came in the same results in which Meta reported $34.91 billion of capital spending for the quarter and negative free cash flow of $2.12 billion.

Are banks banning AI agents?

Not exactly. JPMorgan, Bank of America, Citi and Schwab say outside AI agents can read account data through approved connections, but they block agents that log in with a customer's password and agents that try to move money. No customer losses had been confirmed as of 23 October.

What should I do about the PHP2Shell vulnerability?

If you run a website on PHP, including WordPress, check your PHP version (in WordPress: Tools, Site Health, Info, Server) and update to PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35 or newer, or ask your host to confirm the fix. Updating WordPress or plugins alone does not fix it. Sri Lanka CERT rated the alert Critical.

How do I protect myself from AI deepfake scams?

If a familiar face or voice asks for money, an OTP or bank details, hang up and call back on a number you already have. Never share your bank login, PIN or OTP with anyone, including an AI app. Sri Lanka Police and INTERPOL both issued this advice in October 2026.

Sources

Wins and new tools

DeepSeek's week

Follow the money

Fails

Banks vs AI agents

The PHP flaw

Sri Lanka and the region

Wildcard and breakthroughs

Want next week's episode? Follow Cocoon, and find out where you stand with AI in five minutes.

Take the free AI Readiness Score →