AI & SaaS News, Oct 20: DeepSeek Ban, Meta Muse, Mercury 3

Last checked October 20, 2026 · News roundup

Today in brief (AI news, October 20, 2026): The US government has formally ordered DeepSeek out of federal civilian agencies, and a Reuters investigation shows how one hacker pointed AI agents at more than 1,200 websites for about $5,000. Meta had a rough weekend with AI agents on both sides: its own Muse assistant shared a Marketplace seller's home address, and it blamed a Claude-linked tool for suspending a protest group's Instagram account. On the business side, Canadian publishers are suing Anthropic, a Windows Server update is breaking Outlook on the web for on-premises Exchange, and Town and Tenstorrent closed big rounds.

Most of today's items need no action from regular users. Three do: federal contractors that resell DeepSeek-based software, Exchange admins, and anyone who lets an AI agent act from their accounts. The board below sorts the stories by who has to do something.

Part 1 · AI

Is DeepSeek banned in the US? Federal agencies must now remove it

Not for consumers, but federal civilian agencies can no longer buy or run it. The Department of Homeland Security filed a notice with the Federal Register on October 19, published October 20, that bars IT products and services from Hangzhou DeepSeek Artificial Intelligence and its affiliates from civilian agency procurements. The same order tells agencies to remove DeepSeek products already in place, including third-party products and services built on DeepSeek implementations.

DHS used its powers under the Federal Acquisition Supply Chain Security Act of 2018. CyberScoop reports that the underlying order letter from Homeland Security Secretary Markwayne Mullin is dated June 26, and that an agency wanting an exception has to ask 60 days ahead with a "compelling justification," such as a mission-critical need. DHS says DeepSeek presents "multiple threat vectors," pointing to Chinese laws on government data access and to user prompts that may expose data.

Who it affects: civilian agencies, and any software vendor whose product sold to them runs on a DeepSeek model. The public notice does not list a removal deadline. If you sell to the federal government and use DeepSeek's open weights anywhere in your stack, document it now and plan a model swap. Private companies and individuals are outside this order.

Meta Muse on Facebook Marketplace: the agent shared a seller's home address

Meta's Muse agent, which launched in September, accepted an offer on a Toronto creator's Facebook Marketplace listing and gave the buyer his home address, according to a Fortune report published October 19. The seller, Matt Robb, had let Muse handle the sale of a keyboard as an experiment in late September. The buyer showed up late at night within a day, and Muse sent "Yep, I'm here!" from Robb's account without him knowing.

On another listing, Muse appeared to offer a price $100 below the minimum Robb had set. Meta spokesperson Andy Stone acknowledged the pricing error. Robb says the address came from "always allow" permissions he granted when he first connected Muse to Marketplace; he assumed it would check with him first. Meta itself warns that Muse can "make mistakes or take unexpected actions."

What to do: if you use Muse for selling, swap any "always allow" grant for prompts you approve one by one, and keep your address out of the profile details the agent can read. Muse also lets you delete message history. The same advice fits any agent that can message buyers for you.

Why was the No Kings Instagram account suspended? Meta points to a Claude tool

Meta now says a tool linked to Anthropic's Claude triggered the 90-minute suspension of the No Kings Coalition's Instagram account on Saturday, October 17. Andy Stone, Meta's vice president of communications, first called it an error caused by high posting and commenting rates. On Sunday he gave a new explanation: per The Next Web, Meta engineers found the tool read profiles, posts and hashtags at about four times the speed of a person, in short bursts that looked like bot traffic.

Stone did not name the tool or say who ran it, and the report carries no comment from Anthropic. The page came back first; administrators got full access at 5:33 p.m. ET. The timing hurt, because the coalition was running nearly 3,000 early-voting events ahead of the November 3 midterms.

Why it matters to you: if a business runs a browser agent on its own social accounts, for research, replies or scheduling, the platform may read that speed as automation. Slow the agent down, keep it off accounts you cannot afford to lose, and have a second admin ready to appeal.

Canadian publishers sue Anthropic over Claude training

A coalition of Canadian news publishers has sued Anthropic and its Canadian subsidiary in the Ontario Superior Court. The Canadian Press reported the case on October 19; Broadcast Dialogue says the claim was filed October 6. The plaintiffs are The Canadian Press, The Globe and Mail, the Toronto Star, Metroland Media, Postmedia Network and PNI Maritimes.

They allege breach of contract, copyright infringement and unjust enrichment. The claim says ClaudeBot and other crawlers scraped their articles to train and run Claude, got past paywalls and ignored site terms. As an alternative to general damages, the publishers seek statutory damages of up to $20,000 per infringed work, the maximum under Canada's Copyright Act, plus punitive damages and an injunction. None of this has been tested in court, and Anthropic has not filed a defense; defendants served outside Ontario get up to 60 days to respond.

What changes for Claude users: nothing today. The claim also notes that, unlike Anthropic, most large AI companies have licensing deals with publishers, which suggests a settlement could involve a license deal rather than a product change.

Mercury 3 price: $0.20 per million input tokens, half off at launch

Inception released Mercury 3, the newest version of its diffusion language model, on October 19. Diffusion models write a whole block of text at once and refine it, instead of going one token at a time, which is where the speed comes from: Inception reports 1,068 tokens per second on Nvidia GPUs. The context window grows to 1 million tokens, up from 260,000 in Mercury 2.5.

The API is OpenAI-compatible and runs on Inception's own platform, OpenRouter, Baseten and Vercel, and new accounts on Inception's platform get 100 million free tokens. Inception says Mercury 3 beats GPT-6 Luna and Claude Haiku 5.5 on several benchmarks, but most scores sit in an image on its blog, so treat them as vendor claims. Who should look: teams running high-volume, latency-sensitive jobs such as autocomplete, extraction or voice agents.

How one hacker used AI agents to breach 100+ company websites

A Chinese-speaking hacker let AI agents run for hours on their own against e-commerce sites in the US, Canada, Australia and Japan, Reuters reported in an exclusive on October 19. The campaign started September 25 and hit websites belonging to more than 100 companies, including Deere, SharkNinja and Cummins. According to Reuters, it appeared to use DeepSeek and Anthropic's Claude Code, running on Cairn, an open-source agent platform whose author says it "was not designed for hacking attacks."

Snorre Fagerland of NorForge found about two dozen malicious Cairn instances from around October 2. Deere said it saw "no indication that sensitive information was accessed," SharkNinja said only a product-support site was hit, and Cummins is investigating a site run by a vendor. What to do: if you run an online store, check that your storefront plugins are patched and that someone reviews login and admin alerts daily. The attacks are cheap and patient, not clever, so basic hygiene still matters.

Part 2 · SaaS & business

Outlook on the web not working on Exchange? The October Windows update is the cause

If users on an on-premises Exchange server suddenly cannot sign in to Outlook on the web, or their mail sits in the Outbox, the October 2026 Windows security update is the likely culprit. Microsoft says the update, on Windows Server 2019 and later, added a default cap of 512 simultaneous Endpoint Mapper lookup calls, which a busy Exchange server can exceed, BleepingComputer reported on October 19. Admins may also see "MapiExceptionMdbOffline" (ec=1142) and Event ID 1006 in the application log.

Microsoft's temporary fix, per the report:

1. Open an elevated Command Prompt on the Exchange server.
2. Run: reg add "HKLM\SOFTWARE\Microsoft\Rpc" /v EpmapSavedContextCap /t REG_DWORD /d 65536 /f
3. Reboot, then watch the server and raise the limit again if the errors return.

Microsoft says the October Exchange security updates are not the cause and asks admins not to uninstall them. A permanent fix has no date yet. This only hits Exchange servers you run yourself. If mail leaves the Outbox but lands in junk folders, that is a separate problem; our guide on why emails go to spam and how to fix it walks through the usual causes.

Town AI assistant raises $108M: what Town costs

Town, which gives each employee an AI assistant called a "Townie," announced a $108 million Series B on October 19, TechCrunch reports. Forerunner Ventures and Menlo Ventures co-led, with Andreessen Horowitz, Serena Ventures and First Round Capital joining. The round values Town at $1 billion or more (TechCrunch calls it a unicorn but gives no exact figure) and takes total funding past $160 million, including a $55 million Series A in June. Founder Jean-Denis Greze was previously CTO of Plaid.

Townies connect to email, calendars and other work tools, share approved information with each other and now deal with customers outside the company. Here is what the published plans cost today:

The catch is credits: heavy routines burn them, and overage is billed per credit once you raise the cap. If you mostly want contact and company lookups rather than a do-everything assistant, a dedicated tool may cost less; see our picks of the best sales prospecting tools for small teams, or browse the sales outreach and CRM reviews hub.

Tenstorrent raises $1.5 billion: a bigger rival to Nvidia hardware

Tenstorrent, the AI chip company led by Jim Keller, raised a $1.5 billion Series E at a $6.2 billion valuation on October 19, SiliconANGLE reports. Trace Capital and Innovation X co-led, joined by Protagonist, Morgan Stanley, Fidelity, ARK Invest and T. Rowe Price. Its last raise, about $693 million in December 2024, was reported at a $2.6 billion valuation, so the price has more than doubled in under two years.

The money goes into building more Galaxy servers, which link into large clusters for training and running models, plus RISC-V research and hiring. Tenstorrent also sells Blackhole PCIe cards and the liquid-cooled TT-QuietBox desktop workstation. Virtu Financial runs Galaxy servers for on-premises trading agents, and Equinix offers them to enterprise customers. Prices were not disclosed. Why SaaS buyers should care: more non-Nvidia capacity tends to push down inference costs over time, which is what sets the per-token prices you pay to model vendors such as Inception above.

Questions readers are asking on October 20

Can I still use the DeepSeek app in the US?

Yes, for personal and most business use. The October 20 DHS order covers federal civilian agencies and the products they buy, including third-party tools built on DeepSeek. It does not ban the app for the public, although some states already block it on government devices. Companies with federal contracts should check their software for DeepSeek components.

Is Outlook down, or is it my Exchange server?

If your mailbox is hosted in Microsoft 365, this October bug is unlikely to be the cause. It affects on-premises Exchange servers on Windows Server 2019 or later after the October 2026 Windows security update. Admins can apply Microsoft's registry workaround, raising the EpmapSavedContextCap value, and reboot while a permanent fix is prepared.

Can an AI agent share my address without asking?

It can if you granted broad permissions. In the Fortune case, a seller's "always allow" setting let Meta's Muse share his home address with a Marketplace buyer. Review every agent's permissions, prefer one-time approvals for messages and payments, and keep personal details out of profiles an agent can read.

Does the Canadian lawsuit affect Claude subscriptions?

No. The case is at the start, Anthropic has not filed a defense, and no court has ruled on the claims. Plans, prices and features stay the same. The publishers want damages and a ban on using their work without consent, so any later change would most likely be a licensing deal, not a product change.

Sources: DHS, Federal Register notice on DeepSeek (Oct 20, 2026) · CyberScoop, DHS bans DeepSeek (Oct 19, 2026) · Fortune, Meta Muse on Marketplace (Oct 19, 2026) · The Next Web, No Kings suspension (Oct 19, 2026) · The Canadian Press via CTV, publishers sue Anthropic (Oct 19, 2026) · Broadcast Dialogue, claim details (Oct 2026) · Inception, Mercury 3 (Oct 19, 2026) · Reuters via KFGO, AI-agent hacking (Oct 19, 2026) · BleepingComputer, Exchange issues (Oct 19, 2026) · TechCrunch, Town Series B (Oct 19, 2026) · Town pricing (checked Oct 20, 2026) · SiliconANGLE, Tenstorrent Series E (Oct 19, 2026)