AI News Today (Nov 15, 2026): Nvidia Stock Sale, Claude Opus 5.5
Last updated: Nov 15, 2026 — next refresh daily.
AI news today, November 15, 2026: An Nvidia director sold about $946 million in stock before Jensen Huang resigned. Anthropic made Claude Opus 5.5 its default, and research found export-banned Nvidia chips behind China's top open models. Here are the 10 top stories and 5 security stories, each with free, non-paywalled sources.
Key Takeaways
- Nvidia director Mark Stevens sold about $946 million in stock before Huang resigned.
- Anthropic made Claude Opus 5.5 the default on paid plans and told users to drop "think hard."
- Research found export-banned Nvidia Blackwell chips behind China's top open models.
- Microsoft restored services after a global 365 outage and apologized for the delay.
- Cursor said a stolen Fireworks key let attackers ship a malicious Claude Code extension.
1. Nvidia Director Sold $946M Before Huang's Exit
Nvidia director Mark Stevens sold about $946 million in stock in the months before CEO Jensen Huang resigned on November 12. Stevens, the company's second-largest individual shareholder, sold roughly 4.2 million shares in three batches during the third quarter.
The sales began in late August, after Nvidia's strong Q2 earnings, and continued into September. Stevens still holds about 27.5 million shares. The timing has drawn attention because the scandal erased hundreds of billions from Nvidia's market value.
Stevens joined Nvidia's board in 2008 after an earlier stint from 1993 to 2006. His sales were split into three batches, and the first came after Nvidia's Q2 earnings sent the stock up about 10%.
Why it matters: Insider sales are legal and often routine, but a near-billion-dollar exit just before a governance crisis invites scrutiny from shareholders and regulators.
- Source: Nvidia director sold nearly $1B in stock weeks before the Jensen Huang scandal broke — Yahoo Finance
- Source: Q3 2026 insider stock sales: Nvidia's Mark Stevens tops AI sellers — Quartz
2. Anthropic Sets Claude Opus 5.5 as Default
Anthropic made Claude Opus 5.5 the default model on Pro, Max, Ultra and Team plans. It also published a prompting guide that tells users to drop phrases like "think hard" and start at the Medium effort level.
Anthropic says Opus 5.5 always reasons on its own and scales depth to the task. In its tests, Medium matched the previous Opus 5 at High, while costing 40% less and running more than 30% faster. It costs $4 per million input tokens and $20 per million output tokens.
Why it matters: The guidance shows how quickly prompting advice goes stale: techniques that helped older reasoning models can now add latency without improving answers.
- Source: Claude Opus 5.5 — Anthropic
- Source: Anthropic Publishes Prompting Guidance For Claude Opus 5.5 — Search Engine Journal
3. Export-Banned Nvidia Chips Behind Chinese Models
Research from former OpenAI researcher Will DePue found fingerprints of export-banned Nvidia Blackwell chips in the weights of leading Chinese open models. The analysis covered releases from DeepSeek, Moonshot AI, MiniMax and Ant Group.
DePue's group, Unemployment Labs, says six models released since April 2026 contain numerical patterns matching Blackwell, while earlier models match Hopper. The evidence comes from the published weights, not leaks or insider accounts. U.S. officials had already alleged DeepSeek and Moonshot used Blackwell chips without publishing technical evidence.
Why it matters: The findings suggest export controls have leaked at scale, and they give Washington a technical basis for questioning Nvidia's China sales beyond the spy scandal.
- Source: Export-banned Nvidia chips are behind China's most powerful AI models — TechCrunch
- Source: Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs — Unemployment Labs
4. Anthropic CEO Warns of Uncertainty Before IPO
Anthropic CEO Dario Amodei called buying compute the company's "deepest business dilemma" days before its IPO. He said capacity must be purchased one to two years before anyone knows whether revenue will justify the cost.
Anthropic's commitments have reached nearly $901 billion, mostly for compute and chips, with about $71 billion due by September 2027. Amodei said the IPO will give the company more cash buffer and that he wants to "underpromise and overdeliver." The IPO is meant to cover a potential mismatch between spending and revenue.
Why it matters: The largest AI IPO in history is also a bet that demand will arrive before the bills do, and Amodei says the timing is genuinely uncertain.
- Source: Anthropic CEO Amodei admits 'there's a huge uncertainty in what's going to happen next year' days before a record IPO — The Decoder
- Source: Anthropic CEO Dario Amodei aims to 'underpromise and overdeliver' as IPO valuation drops below $2T hopes — Yahoo Finance
5. OpenAI Disbands Standalone Robotics Team
OpenAI disbanded its standalone robotics team and reassigned the researchers to other groups. The company says robotics remains a core part of its long-term vision, but general-purpose models like GPT-6 Astra can now control robots directly.
In an Nvidia study, an Astra-powered agent completed 74.8% of simulated robotics tasks, compared with 51% for the best specialized models. A separate benchmark found Astra finished 7 of 100 real desk-object tasks, while a specialized model finished none. OpenAI had booked a booth at the Conference on Robot Learning, which sat empty after the reorganization.
Why it matters: OpenAI is betting that embodied intelligence will emerge from general models rather than a dedicated robotics stack, a bet that could reshape how robot startups build.
- Source: OpenAI reassigns robotics researchers to other teams as general-purpose models might make specialized robotics AI unnecessary — The Decoder
- Source: OpenAI disbands stand-alone robotics team as general AI models start running robots — Fortune
6. Gemini 4 Carbon Tops MathArena at 15x Cost
Google's Gemini 4 Carbon became the first non-OpenAI model in months to top the MathArena benchmark, scoring 85.4% and knocking GPT-6.1 Sol from first place. Anthropic's Claude Fable 5.2 came second at 82.8%.
The win comes at a price: Carbon cost about $15.70 per problem in testing, roughly 15 times GPT-6.1 Sol's $1.07. Fable 5.2 cost about $40.20. MathArena says the scores reflect the reasoning settings used in its tests. The site added the Carbon and Fable results on November 13.
Why it matters: Benchmark leadership is increasingly a trade-off between accuracy and cost, and a 15x premium makes Carbon a specialist tool rather than a default choice.
- Source: Gemini 4 Carbon takes the MathArena crown but costs 15 times more than GPT-6.1 Sol — The Decoder
- Source: MathArena leaderboard — MathArena
7. Banks Bar Chinese Investors from Anthropic IPO
Banks handling Anthropic's IPO are rejecting orders from mainland China and Hong Kong, citing US export-control rules known as ITAR. Some Chinese investors are trying offshore routes, including a British Virgin Islands trust tied to a tech founder.
Goldman Sachs and Morgan Stanley lead a syndicate of 22 banks. The shares are set to list on Nasdaq under the ticker ANT after pricing on November 18, with trading expected to begin the next day.
Why it matters: Export controls are now shaping who can buy a US AI IPO, turning a capital-markets event into another front in the technology competition with China.
- Source: Anthropic IPO Reportedly Closed To China, Hong Kong Buyers, Mirroring SpaceX Playbook — Yahoo Finance
- Source: Banks rebuff Chinese and Hong Kong orders for Anthropic's IPO over US arms-export rules — MLQ.ai
8. Runway Open-Sources Praxis-1 Robot Model
Runway released Praxis-1, an open-weight robot control model, with weights and source code under the Apache 2.0 license. The model uses video pretraining plus robot data to turn camera images, text instructions and position data into movements.
Praxis-1 scores 46.2% on Nvidia's RoboLab-120 benchmark, topping the World Action Model category. Runway says the model can be adapted to new robots with additional training, though it still confuses similar objects and cannot plan complex sequences reliably. Runway first introduced Praxis-1 in September and released the weights on November 11.
Why it matters: Open weights let robotics startups fine-tune a capable controller without building a foundation model, but the model's limits show how far video-trained control still has to go.
- Source: Runway's open-source Praxis-1 turns video AI into a robot controller — The Decoder
- Source: Introducing Praxis-1: Runway's Open-Weight World Action Model — Runway
9. Microsoft Apologizes for Global 365 Outage
Microsoft apologized after a global outage knocked out Outlook, Teams and Copilot for nearly 24 hours. The company said the disruption began during a network capacity upgrade and was not caused by a cyberattack or an AI agent.
The outage affected users in more than 70 countries and disrupted businesses, courts, hospitals and government services. Microsoft said on Saturday that service had been fully restored and that it had addressed the underlying issue. The disruption began shortly before 1 p.m. ET on Thursday.
Why it matters: The incident is a reminder that a small number of cloud providers now underpin critical services, and a bad configuration can rival a cyberattack in impact.
- Source: Microsoft apologizes for massive outage that took down Outlook and other 365 services — CNBC
- Source: Microsoft says services fully restored after day-long global outage — BBC
10. Nvidia Investors Urge Board to Consider Huang's Return
Altimeter Capital's Brad Gerstner and Atreides Management's Gavin Baker asked Nvidia's board to run a slow, deliberate CEO search that leaves open the possibility of Jensen Huang's return. The firms disclosed about $2.13 billion of Nvidia stock combined as of June 30.
Huang resigned as president, CEO and director on November 12 after reports about personal relationships and amid an independent board investigation. Nvidia's resignation announcement did not state a reason. The investors said Huang's long record and the uncertainty around his departure justified a slower process.
Why it matters: The letter shows some major shareholders are not ready to close the door on Huang, even as the board investigates the conduct that forced him out.
- Source: Nvidia board urged to consider 'potential return of Jensen Huang' by two major investors — CNBC
- Source: 'Don't rush to replace Jensen Huang': Nvidia investors urge board to keep comeback option open — CNBC TV18
AI Security: 5 Stories to Know Today
11. Cursor Breach Led to Malicious Claude Code Extension
Cursor confirmed that attackers used an access key stolen from AI infrastructure provider Fireworks AI to publish malicious versions of Anthropic's Claude Code extension in its marketplace. The compromised versions, 2.1.312 and 2.1.313, tried to install a persistent SSH backdoor.
Cursor says up to 5,000 developers were exposed. The stolen key also let attackers copy historical coding-agent traces containing unredacted credentials, which were used to compromise additional services. Cursor published remediation steps and indicators of compromise. The malicious versions were available for about 11 hours.
Why it matters: AI coding tools sit on developer machines with broad access, and one leaked key in the model supply chain can turn an update into a backdoor.
- Source: Security notice — malicious version of Anthropic.claude-code (2.1.312 and 2.1.313) identified in the Cursor VS Code Extension Marketplace — Cursor
- Source: Cursor confirms security breach led to malicious Claude Code extension — BleepingComputer
12. ArcticRAT Uses AI to Profile Victims' Finances
Zimperium found that ArcticRAT, an Android remote access trojan, now sends stolen SMS messages and app data to xAI's Grok model. The AI sorts bank alerts, one-time codes and transaction notifications from routine messages.
Researchers identified 78 unique samples, 42 droppers, 36 payloads and 21 command-and-control IP addresses. The malware can also take full control of a device through hidden remote access and a black-screen overlay, and it estimates a victim's financial standing before exfiltrating data. Early versions targeted users of 11 cryptocurrency apps before expanding.
Why it matters: AI is turning malware from a data vacuum into a triage system, letting criminals find the most valuable victims and accounts without manually reading every stolen message.
- Source: Uncovering ArcticRAT: The Evolution from Targeted Crypto-Stealer to AI-Powered Android RAT — Zimperium
- Source: ArcticRAT Cybercrime Gang Profiles Victim Finances With AI — Dark Reading
13. FBI Warns Artex Automates Attacks at Machine Speed
The FBI released a flash alert on Artex, an open-source AI platform that can turn a large language model into an autonomous hacking tool. The agency says Artex automates reconnaissance, exploitation and data exfiltration at machine speed.
Some intrusions progressed from initial targeting to privileged access in under three hours, the FBI said. The tool was released on GitHub in August and remains available through community forks even after the original repository was removed. The FBI said it has spread through localized variations and archived copies.
Why it matters: Agentic attack frameworks lower the skill barrier and compress the time defenders have to respond, turning what used to be a weeks-long intrusion into an afternoon.
- Source: FBI flash alert says Artex tool automating cyberattacks at 'machine speed' — The Record
- Source: Threat Overview of ARTEX Autonomous Penetration Testing Framework — FBI IC3
14. N-able Zero-Day Exploited on 21 N-central Servers
Attackers exploited a new N-able N-central zero-day called 4EverAble2Shell within hours of a public proof of concept. Researchers found 21 compromised servers, including 19 running N-able's latest Hotfix 2 release.
The attackers deployed an XMRig Monero miner configured to use about half the available CPU cores. N-able restricted some N-central features and integrations and urged customers to disable legacy API access and upgrade agents and probes. The compromised servers were spread across the US, Canada, the UK, Germany, France, Sweden and South Africa.
Why it matters: An RMM platform gives attackers root on a host that manages thousands of customer endpoints, so a single compromise can cascade across managed service providers.
- Source: Hackers Exploit N-central Flaws to Infect 21 Servers, Including 19 Running Hotfix 2 — The Hacker News
- Source: New N-able '4EverAble2Shell' zero-day exploited hours after PoC released — BleepingComputer
15. CISA Flags Exploited UnrealIRCd Command Injection
CISA added CVE-2026-114713, a command-injection flaw in UnrealIRCd, to its Known Exploited Vulnerabilities catalog. The bug lets any user on a network with two or more linked servers inject server-to-server commands executed with server authority.
UnrealIRCd fixed the flaw in version 6.2.7.1, and exploitation attempts began a day later. The project's maintainer said attackers were exploiting the issue on the internet and urged administrators to update or deploy the spamfilter workaround. The flaw affects UnrealIRCd 6.0.0 before 6.2.7.1.
Why it matters: IRC servers are often forgotten infrastructure, and a command-injection flaw with server-level privileges can hand attackers a durable foothold.
- Source: Attackers Exploit UnrealIRCd Flaw That Enables Server-Level Command Injection — The Hacker News
- Source: CISA Adds Two Known Exploited Vulnerabilities to Catalog — CISA
More AI Stories Worth Reading
- Elon Musk and other tech executives signed a petition with more than 115,000 signatures calling on Nvidia to reinstate Jensen Huang. — TechCrunch
- AI tools are helping less sophisticated criminals launch more ransomware attacks, according to Wired. — Wired
- Young job seekers describe AI-filtered applications and a broken entry-level job market in a new Guardian series. — The Guardian
Related Reading on Kill The AI
- Nvidia's insider sale, Anthropic's IPO demand, and ChatGPT Wallet.
- Huang's resignation, Gemini 4 Hydrogen, and Nebius's $40 billion plan.
- OpenAI's $10 billion quarter, TSMC's AI bet, and AlphaGenome 2.
- Our data review of the AI boom's first eight weeks, from capex to layoffs.
Methodology & Sources
Compiled November 15, 2026, using multi-source research. Outlets: Yahoo Finance, Quartz, Anthropic, Search Engine Journal, TechCrunch, Unemployment Labs, The Decoder, Fortune, MathArena, MLQ.ai, Runway, CNBC and BBC. Also CNBC TV18, Cursor, BleepingComputer, Zimperium, Dark Reading, The Record, FBI IC3, The Hacker News, CISA, Wired and The Guardian. All sources are free to read; no paywalled links. Details are as reported at compilation time.
Frequently asked questions
Nvidia director Mark Stevens sold about $946 million in stock before Jensen Huang resigned. Anthropic made Claude Opus 5.5 its default, research found export-banned Nvidia chips behind China's top open models, and Microsoft apologized for a global outage. The FBI also warned that the Artex AI attack tool is automating intrusions.
Nvidia director Mark Stevens sold about $946 million in company stock, or roughly 4.2 million shares, in three batches during the third quarter. The sales began in late August, before Jensen Huang resigned on November 12. Stevens remains Nvidia's second-largest individual shareholder and still holds about 27.5 million shares.
Anthropic made Claude Opus 5.5 the default model on Pro, Max, Ultra and Team plans. It published a guide telling users to drop phrases like 'think hard' and start at the Medium effort level. Anthropic says Medium matches the previous Opus 5 at High while costing 40% less and running more than 30% faster.
Former OpenAI researcher Will DePue analyzed the published weights of models from DeepSeek, Moonshot AI, MiniMax and Ant Group. He found that six models released since April 2026 contain numerical patterns matching export-banned Nvidia Blackwell chips, while earlier models match Hopper. The evidence comes from the model weights, not leaks.
A global Microsoft outage knocked out Outlook, Teams and Copilot for nearly 24 hours and affected users in more than 70 countries. Microsoft said the disruption began during a network capacity upgrade and was not caused by a cyberattack or an AI agent. The company apologized and restored service on Saturday.
Last updated: Nov 15, 2026 — next refresh daily. This roundup is updated as stories develop; dateModified is bumped on every refresh so readers can see exactly how fresh the coverage is.