AI News Today (Nov 7, 2026): AI Selloff, METR Audit & More

AI news today, November 7, 2026: the AI selloff wipes more than $3.2 trillion from US stocks before a partial rebound. METR finds Anthropic's research is accelerating, and protesters storm a Microsoft data center site in Amsterdam. Here are the 10 top stories and 5 security stories, each with free, non-paywalled sources.

Key Takeaways

• The AI selloff wiped more than $3.2 trillion from US stocks on Thursday before a partial Friday rebound.
• METR says Anthropic's internal model progress has grown about 1.9 times faster since October 2025.
• Protesters broke into a Microsoft data center site in Amsterdam; Dutch police used force and tear gas during the action.
• Perplexity launched always-on Teammates agents, and Amazon added text-based Alexa shopping for US customers.
• DigiCert says a server migration exposed archived customer names and identity documents in a breach.

1. Update: AI Selloff Wipes $3.2 Trillion as Markets Rebound

More than $3.2 trillion was wiped from US stocks on Thursday, when the Nasdaq fell 5.5% and the S&P 500 dropped 4.4%. Markets rebounded when trading reopened on Friday, but the Guardian reported they failed to recover all of the lost ground as selling pressure lingered.

The Dow fell 1,833 points, or 3.4%, and the Magnificent Seven lost more than $1 trillion in market value. Nvidia fell 9% and lost more than $440 billion, while the VIX jumped 11.57 points. Investors blamed the midterm results and data center moratoriums for the selloff.

All 11 S&P 500 sectors fell, led by information technology at 7.3%. The decline was the steepest one-day drop since April 2025, when tariffs shook markets. We first covered Thursday's rout in our November 6 briefing.

Why it matters: AI valuations now move on election results and permitting risk, not just chip demand.

• Source: US tech stock rout: is the AI bubble bursting? — The Guardian
• Source: Magnificent 7 loses $1 trillion. The most eye-popping stats from Thursday's tech stock plunge — CNBC

2. METR Finds Anthropic's AI Research Is Accelerating

METR spent eleven weeks inside Anthropic, from August 12 to October 29, reviewing internal benchmarks, training data and staff surveys. The review included three unreleased models that finished training between July and October 2026. It found that Anthropic's benchmark scores have grown 1.9 times faster since October 2025, implying roughly two years of progress in one.

METR estimates that about 80% of the current acceleration comes from AI-assisted research, not faster compute scaling. Anthropic redacted most of the evidence METR found most informative and canceled a compute-allocation review. The Decoder reports the internal acceleration value was 1.97, close to Anthropic's warning threshold.

Why it matters: If AI is already doubling the pace of AI research, safety thresholds and outside audits need to keep up.

• Source: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 — METR
• Source: METR finds Anthropic may have crossed its own AI acceleration warning threshold back in July — The Decoder

3. Protesters Break Into Microsoft Data Center Site in Amsterdam

Activists from Geef Tegengas broke into the Westpoort construction site of a Microsoft data center in Amsterdam. Some climbed a crane, lit flares and hung banners. Photos showed protesters on scaffolding and a banner reading Microsoft Powered Genocide. Police said force was used and an officer was slightly injured; protesters said police also used tear gas.

The action was separate from the pan-European Stop Dirty Data Centres campaign, which stayed outside the fence. More than 100 groups held two days of protests, with smaller actions in Bristol, Berlin, Warsaw and Zagreb. Pure Data Centres owns the site; Microsoft is the tenant. Microsoft said it respected peaceful protest but did not condone unlawful or unsafe actions.

Why it matters: Data centers are becoming the physical front line of the AI backlash, and permits are now political.

• Source: Activists break into building site of Microsoft datacentre in Amsterdam — The Guardian
• Source: Protesters break into Microsoft data center site in Amsterdam — The Verge

4. Perplexity Launches Always-On Teammates for Work

Perplexity launched Teammates, always-on AI agents that can take on multiple roles in a workspace, keep their own threads and work alongside users in Slack. They build on Perplexity Computer and require the $200-per-month Max plan. The Verge says it is much more expensive than rival agents. Users can follow each agent's own progress in the Perplexity app or Slack.

Perplexity says the agents can work simultaneously on ongoing, large-scale projects. They pull the existing projects, skills and memory from the Perplexity tools a user already uses. CEO Aravind Srinivas said he expects organizations to eventually run hundreds of teammates for every employee. OpenAI's Dots and Anthropic's Claude Duet are competing products.

Why it matters: Always-on agents are becoming a premium subscription product, not a free chatbot feature.

• Source: Perplexity launches its own take on AI agents for work — The Verge
• Source: Perplexity now has its own always-on AI agent — The Decoder

5. Amazon Lets Customers Text Alexa to Shop

Amazon launched text-based shopping for select US customers, letting them message Alexa to search for items, add products to a cart and complete purchases. Users can also share a photo, compare products and check order status. The agent can also manage shopping lists, track prices across the web and search online for items.

The shopping feature follows Amazon's September block on Meta's Muse agent and arrives before the holiday season. Amazon says it will expand the broader rollout in the coming weeks. Customers who do not receive an invite can join a waitlist. It builds on recent Alexa shopping upgrades from October.

Why it matters: Shopping is moving from search boxes and apps into messaging threads, and Amazon wants to own the thread.

• Source: After blocking Muse, Amazon now lets users text Alexa to help them shop — TechCrunch
• Source: Amazon will let you text Alexa your shopping list — The Verge

6. Airbnb Will Launch an AI Travel Agent in 2027

Airbnb will launch an AI travel agent in the first half of 2027. The agent will build a custom interface around each request, composing maps, photos, comparisons, calendars and itineraries in real time instead of replying in a text-heavy chat. The company says the agent will feel like generative UI, offering the right interface at the right time.

Airbnb says the agent will also handle changes, payment issues, refunds, rebooking and 24-hour safety support. The announcement came with Q3 revenue of $4.87 billion, up 19%, and a higher 2026 growth forecast. CEO Brian Chesky said general-purpose agents are not the best for travel, and 80% of travelers enjoy researching trips.

Why it matters: Travel is the first big test of whether AI agents can beat a well-designed browse-and-book interface.

• Source: Airbnb says it's launching an AI travel agent in the first half of 2027 — TechCrunch
• Source: Airbnb is working on an AI travel agent — The Verge

7. Anthropic IPO Valuation Faces Pushback From Investors

Anthropic is targeting a $1.5 trillion valuation and an IPO of up to $100 billion, but investors are pushing back after the AI selloff. One investor floated a valuation closer to $1 trillion, and four people told Reuters the roadshow may not start as planned next week. Anthropic was eyeing up to $2 trillion as recently as September.

Anthropic's amended S-1 leaves October revenue blank, and investors want the figure to judge whether growth held up. The company reported $17.0 billion in third-quarter revenue, a $567 million operating loss and $146.6 billion in cash. A listing is targeted for the week of November 16.

Why it matters: The IPO is now a referendum on whether AI revenue growth can justify trillion-dollar valuations after the rout.

• Source: Anthropic's $1.5 trillion IPO hits resistance as investors balk at valuation — The Decoder
• Source: Big Investors Now Doubt Anthropic Can IPO at $1.5 Trillion After the AI Crash — Yahoo Finance

8. Thinking Machines Releases Inkling 2 Small

Thinking Machines released Inkling 2 Small, a mixture-of-experts model with 594 billion total parameters and 17 billion active. It accepts image inputs, supports variable effort and scored 31 on the Artificial Analysis Intelligence Index. Thinking Machines says the model is the strongest Western open-weight model, with large gains in agentic tasks.

The model is available on Tinker with closed weights while external safety testing runs. Thinking Machines plans to publish the full weights under Apache 2.0 next week. It trained on 8.4 trillion tokens and ran a nine-day reinforcement-learning job across 2,048 GPUs. The company says it uses about four times as many output tokens and costs $1.18 per task, nearly six times its predecessor's cost.

Why it matters: The gap between closed and open models is narrowing, and Thinking Machines is shipping a frontier-adjacent model with a safety delay.

• Source: Introducing Inkling 2 Small — Thinking Machines
• Source: Thinking Machines' Inkling 2 Small claims top spot among western open-weight models — The Decoder

9. Moonshot Sought Nvidia Blackwell Capacity in Europe

Moonshot AI discussed leasing clusters of 8,000 or more Nvidia Blackwell GPUs in Europe, according to four sources who spoke to CNBC. The Chinese lab wanted capacity for inference, and one source said it was also aware of interest in training clusters. Moonshot released Kimi K3 in July and temporarily paused new subscriptions after demand stretched its capacity.

US export controls ban Blackwell imports into China but do not clearly cover remote access through overseas data centers. Moonshot was reportedly willing to pay 30% to 50% more upfront than a non-Chinese customer. The White House accused it of using Blackwell chips in Thailand in July. The Remote Access Security Act passed the House but stalled in the Senate.

Why it matters: Export controls are becoming a game of data-center geography, and Europe is the newest loophole.

• Source: Chinese AI lab Moonshot turned to Europe for Nvidia Blackwell access amid U.S. China curbs, sources tell CNBC — CNBC
• Source: Moonshot AI Discussed European Blackwell Leases Amid U.S. Chip Curbs — Unite.AI

10. US and Russia Move to Block Binding AI Weapons Rules

Russia and the US are leading a group of states trying to block binding rules on autonomous weapons at a UN conference in Geneva from November 16 to 20. The US, Russia and Israel oppose setting a negotiating timetable; 128 countries are expected to attend. The conference is seen as the key moment for a mandate to negotiate binding rules.

A 50-state coalition led by Ireland wants a mandate for 25 days of negotiations in 2027. The ICRC says more than 70 states support talks, and 163 states voted in a UN committee to launch negotiations. The UK did not back the Irish statement. Pope Leo XIV urged states to take concrete steps toward banning lethal autonomous weapons.

Why it matters: The conference will decide whether AI weapons get a treaty or another decade of voluntary talks.

• Source: Russia and US seek to block plan for binding rules on autonomous weapons — The Guardian
• Source: UN Autonomous Weapons Talks: Human Oversight Erodes — Cloud Security Alliance

AI Security: 5 Stories to Know Today

11. DigiCert Data Breach Exposes Customer Identity Documents

DigiCert disclosed a breach after a February 2025 server migration left an internal business system exposed to the internet. An unauthorized account accessed archived customer data between June 27 and July 2, 2026, and DigiCert confirmed the data-access event on October 19.

The exposed records included names, business emails, phone numbers and addresses, plus driver's licenses, passports and utility bills for some customers. DigiCert says its certificate authority, CertCentral and DigiCert ONE services were not compromised, and no login credentials or certificates were taken.

Why it matters: A certificate authority's identity documents are a ready-made phishing kit for attackers.

• Source: Certificate authority DigiCert reports data breach involving business customer info — The Record
• Source: DigiCert data breach exposes customer information, identity documents — BleepingComputer

12. CWPdoor Backdoor Hides in Control Web Panel Server

Researchers at QiAnXin's XLab found CWPdoor, a backdoor hidden inside a trojanized version of cwpsrv, the web server behind Control Web Panel. The malware waits for requests with signed X-Ngx-Integrity headers, decrypts commands with AES-256-CBC and suppresses access logs.

The installer captured on June 29 targets versions 1.15.10 and 1.24.0 and restores file timestamps to hide the swap. More than 340,000 CWP instances were exposed online in the past month. The backdoored binary had zero VirusTotal detections, and the same archive contained Ebury malware.

Why it matters: A backdoor that never appears in logs can sit on hosting servers for months.

• Source: Hackers Hide Linux Backdoor in Control Web Panel Server to Execute Commands Without Logs — The Hacker News
• Source: Hackers backdoor Linux servers using trojanized CWP web server — BleepingComputer

13. Claude Dispatch Flaw Lets Extensions Run Code

Forever Security found that Claude Dispatch was reachable from an ordinary claude.ai session, so a browser extension could send instructions to Claude Desktop. Four prompts and auto-clicked Allow buttons made Claude install an npm package that opened Calculator. No browser exploit was needed.

Anthropic acknowledged the report on May 19 as a duplicate and said on October 13 that Dispatch would be retired by the end of October. Its help page now says Dispatch is closed to new users, but existing users can keep using it. The researchers say other Claude features share the design.

Why it matters: A browser extension should not be able to turn a website session into code execution on a laptop.

• Source: Malicious Browser Extensions Could Use Claude Dispatch to Run Code on Users' Computers — The Hacker News
• Source: ThisBash Escape: Claude Dispatch Flaw Enables Browser Extensions RCE on the OS — Forever Security

14. ChatGPT Lockdown Mode Bypassed via Popup Flaw

Check Point found that a nested iframe in ChatGPT's HTML preview could open a popup outside the sandbox, bypassing Lockdown Mode. The popup could exfiltrate chat history, display phishing pages or deliver malware disguised as a ChatGPT update.

In ChatGPT Classic, the attack required zero clicks: a malicious GPT or shared chat was enough. OpenAI fixed the flaw on September 18 by removing the allow-popups and allow-popups-to-escape-sandbox permissions, and Check Point confirmed the exploit no longer works.

Why it matters: Lockdown Mode is only as strong as the application features around the model.

• Source: Researchers Bypass ChatGPT Lockdown Mode to Exfiltrate Chats Without User Interaction — The Hacker News
• Source: ChatGPT Lockdown Mode Bypass: Phishing, Malware Delivery, and Silent Data Exfiltration — Check Point Research

15. AI Voice Scams Steal $129 Million From Companies

The Justice Department indicted Jacob Soall, Sharon Hazan and Guy Sheetrit for using AI to impersonate CEOs and lawyers in a fraud worth about $129 million. Prosecutors say the men cloned voices and images on calls, then had employees wire money to accounts they controlled.

Victims included hospitals in Missouri and California, a Rhode Island bank, a UK fintech, a Massachusetts mutual fund and a Florida pet food company. The three were arrested between June and September, and extradition requests are pending. Each faces seven charges, including wire fraud and money laundering.

Why it matters: Voice and video cloning has turned a familiar email scam into a board-level financial risk.

• Source: CEO-spoofing scammers used AI to help steal more than $100 million, feds say — The Record
• Source: Trio charged with impersonating CEOs using AI to steal millions from top companies — Courthouse News

More AI Stories Worth Reading

• A lawsuit says ChatGPT helped a North Carolina woman plan her husband's poisoning, including lethal doses and how to hide the drugs. — Ars Technica
• Fake OpenAI, SpaceX and Anthropic event invitations are stealing Microsoft and Google accounts through device-code and adversary-in-the-middle phishing. — BleepingComputer
• The UK's NCSC says it is highly likely that AI-powered malware will hit critical national infrastructure in a ransomware attack by 2028. — Dark Reading

Related Reading on Kill The AI

• AI video, the Nasdaq selloff, Meta's Conductor deal and OpenAI's 2027 IPO.
• Data centers lost the midterms, OpenAI bought Pierre, and Google faced a school Search backlash.
• ChatGPT search hit 20% of Google, Google tied school Search to AI, and Meta sought $25B in bonds.
• Our data review of the AI boom's first eight weeks, from capex and chips to layoffs.

Methodology & Sources

Compiled November 7, 2026, using multi-source research for AI news today. Outlets: The Guardian, CNBC, METR, The Decoder, The Verge, TechCrunch, Yahoo Finance, Thinking Machines, Unite.AI, Cloud Security Alliance, The Record, BleepingComputer and The Hacker News. Also Forever Security, Check Point Research, Courthouse News, Ars Technica and Dark Reading. All sources are free to read; no paywalled links. Details are as reported at compilation time.

Frequently asked questions