AI News Today

Top story: Anthropic Shares Fall ~5% After-Hours on Debut as Apple Job Listings Reveal Rival AI-Agent Push · bloomberg.com

The top AI stories and live updates for Friday, November 20, 2026 — selected by the team behind 600+ issues, tracked across 113 entities.
● LIVE Updated 2m ago · Edited by Alexis · Daily editions · About the index

Top AI Stories Today

GAO: ChatGPT, Claude, Gemini, Grok all planned terror attacks

The Government Accountability Office reports that ChatGPT, Claude, Gemini and Grok each produced viable US terrorist attack plans in under 12 hours after investigators bypassed each platform's safeguards. Two models produced IED plans aimed at mass casualties; two others produced plans targeting …

nextgov.com · 22h ago · Field

Snap ships CLI that builds AR lenses via Claude Code and Cursor

Snap launched Easy Lens CLI on Nov 18, letting developers pipe plain-language prompts through Claude Code, Cursor or Codex to build and publish Snapchat AR Lenses end-to-end. The tool runs on Mac, Windows and Linux and plugs into Snap's Lens+ Payouts revenue program, extending the company's bet t…

techcrunch.com · 16h ago · Builders · our brief →

House panel: Commerce OK'd 879,000 AI chip exports to China

House Foreign Affairs Chair Brian Mast on Nov 19 released Commerce export-license data showing BIS approved more than 879,000 Nvidia H200 and equivalent AMD AI chips for sale to China between December 2025 and May 2026, out of 3.4 million requested. Named recipients include Alibaba, Tencent, Byte…

cnbc.com · 6h ago · Geopolitics · our brief →

Kuaishou's Kling Image 4.0 generates native 8K from one prompt

Kuaishou unveiled Kling Image 4.0 on Nov 18 with native 8K image generation, up to 10 coordinated images per prompt, and 2,000 characters of in-image text across seven languages and 70+ commercial fonts. The model is available via API to developers and enterprise customers, targeting advertising,…

technode.com · 16h ago · Builders · our brief →

Kokoro TTS HF account hijacked via lapsed domain, weights edited

Researcher Alon Hertz's team registered the lapsed hexgrad.com domain, used it to reset the Hugging Face account publishing Kokoro (a TTS model with ~13 million monthly downloads), and demonstrated they could fine-tune its weights in 87 seconds on a gaming GPU — turning, in one example, a cake re…

techtimes.com · 12h ago · Field
AI News Pulse
Most covered NVIDIA — in 5 of the last 20 issues · 89 tracked stories this week
Fastest riser Jensen Huang — ▲ +600% story volume vs last week (49 vs 7 tracked stories)
Story volume 416 tracked stories this week vs last week

Latest AI News — Last 48 Hours

Anthropic drops 5% after-hours as Apple preps rival AI agent
Anthropic shares dropped roughly 5% in post-market trading on November 19 after Apple job listings surfaced detailing an in-house autonomous AI agent that writes code, browses the web, controls apps and coordinates sub-agents — directly targeting the agent market Anthropic, OpenAI and Meta are racing to own. The slide pared Anthropic's 17-34% debut-day pop from its record $107.3 billion IPO, with investors questioning whether frontier-lab returns can clear the compute commitments already disclosed. The Nasdaq fell 2.2% on the same session.
Pentagon CIO Kirsten Davies named acting US cyber director
The Trump administration on November 19 tapped Pentagon CIO Kirsten Davies as acting National Cyber Director, replacing Sean Cairncross days after a wave of tech-related outages hit Microsoft, the FAA, banks, United Airlines, Apple and an NSA network. Davies, who previously led global cyber resilience at Unilever and worked at Estée Lauder, Barclays-Africa, HP and Siemens, brings more operational cyber experience than Cairncross, an attorney chief of staff Stephen Miller pushed to exit. The transition lands as the White House also weighs a 12-month liability pause for companies that report cyber incidents.
California AG subpoenas Nvidia over Huang China chip transfers
California Attorney General Rob Bonta on November 19 opened a state investigation into Nvidia and subpoenaed the company, seeking to determine whether its advanced chips or confidential information were routed to China after reports that former CEO Jensen Huang had an affair with a suspected Chinese spy. Bonta said the probe will also examine whether Nvidia misled investors and the public, warning that 'companies cannot mislead their investors' despite the 'messy, transnational situation.' Huang resigned last week; CFO Colette Kress is serving as interim CEO and the SEC's parallel civil probe remains open.
Menlo and Lightspeed emerge as Anthropic's biggest VC winners
Reuters reports Menlo Ventures put $1B+ into Anthropic for a ~1.5% stake now worth $20B+ at the $1.5T IPO valuation, while top financial investor Lightspeed invested $3B+ for a 2%+ position. The disclosures round out the Anthropic VC payout picture after Iconiq's 1.8% stake was reported at $31.8B earlier in the day.
39 states fight Disney's $20M Character.AI IP deal
A coalition of 39 state attorneys general plus DC asked a Delaware bankruptcy judge Nov 19 to block Character.AI's proposed sale of intellectual property to Disney for $20 million, arguing the deal terms unfairly exclude competing bids between $15.2M and $21.7M that would net the estate more. The states also want a consumer-privacy ombudsman appointed before any sale of chat logs from 2.1M under-18 US accounts. Bid deadline is Nov 20; auction Nov 23.
Court: 4th Amendment bars AI-only probable cause
A federal court in Kentucky denied qualified immunity to Florence Police detective Ryan Kilmer, holding that the Fourth Amendment does not let police arrest someone based solely on an uncorroborated AI facial-recognition match. Alain Osorio spent more than two weeks in jail after a department tool matched a bank CCTV screenshot to a Facebook profile. Judge S. Chad Meredith wrote the Constitution 'does not permit an individual to be arrested based on nothing more than an uncorroborated conclusion by AI that the individual in question has committed a crime.'
DeepSeek closes $15.8B round; Tencent and CATL now top holders
Hangzhou DeepSeek completed its industrial-and-commercial change Nov 19, finalizing a ~105.5B yuan ($15.8B) second funding round that brings total raised to roughly 156.5B yuan ($23.5B). Tencent and CATL tied as largest investors at 9.5653% each, with China Mobile and CITIC Securities Investment at 7.65%; the 31-partner cap table adds NetEase, Meituan, Xiaohongshu, vivo, Midea, Geely, IDG and Boyu, while founder Liang Wenfeng was removed from the direct partnership and now controls through a 99%-owned consulting entity. CITIC Securities is engaged for a potential Sci-Tech Innovation Board listing.
Iconiq's Anthropic stake worth $31.8B after day-one pop
Forbes reports Iconiq Capital holds an estimated 1.8% of Anthropic after investing more than $6B, a stake worth about $31.8B at Thursday's $1.76T market close — a 430%+ return and Iconiq's largest ever ownership position, built off a January 2023 cold email to Dario Amodei. Reuters separately pegs Menlo Ventures at a 1.5% stake worth $20B+ on a $1B check, with Lightspeed holding 2%+ off a $3B+ investment.
Superhuman puts Go assistant across Docs and Mail
Superhuman expanded its Go assistant across Docs and Mail with proactive scheduling, inline cursor help, enterprise search across connected tools, and routines that trigger workflows on events. CEO Shishir Mehrotra said Go has over 1 million active users, and the Go SDK now covers 20 new connected agents including DeepL, Legora, Amplitude, Databricks, and Blackboard.
Fireworks AI: open models trail frontier labs by months, not years
On Anthropic's IPO day, Fireworks AI co-founder Chenyu Zhao told TechCrunch that open-weight models trail frontier models by 'at most six months, probably closer to three,' and companies spending $10K-$100K/month on tokens past 10,000 users should 'think about open source.' Fireworks, last valued at $17.5B, is pushing a 'two-model world' thesis, though the company is also recovering from two fall 2026 breaches, including one that exposed 150,000 Uber drivers' license data.
DeepMind essay: AGI-to-ASI inside a decade
DeepMind research scientist Tim Genewein published 'What Comes After AGI?' arguing superintelligence is not a speculative leap but a near-term trajectory, built on a technical report co-authored with 11 DeepMind and academic researchers. The essay identifies four pathways that can take systems past human-level AGI without further breakthroughs: scaling, paradigm shifts, better agent coordination, and recursive self-improvement — with 'sheer scale' alone capable of bridging the gap.
CISA nominee LoSapio ducks priorities, denies Jan 6 was an attack
Trump's CISA director nominee Michael LoSapio gave vague answers on agency priorities, refused to defend the $385M proposed cut, and told Sen. Blumenthal 'I don't believe it was an attack' when asked about January 6, prompting multiple senators to signal they will block his confirmation. LoSapio also declined to say Biden won the 2020 election — a notable stance for the role overseeing election security.
Nvidia nears $500M+ deal for diffusion-LLM startup Inception
Nvidia is in advanced talks to acquire diffusion-based LLM developer Inception for more than $500 million, plus performance-based earnouts tied to new-employee targets, The Information reported Wednesday. Talks began before Jensen Huang's resignation last week. Inception's diffusion architecture is expected to fold into Nvidia's open-source Nemotron project, targeting faster inference than the autoregressive models powering ChatGPT and Claude.
Intuit sells Mailchimp for $3.68B to focus on AI-native SMB tools
Intuit announced Nov 19 it will sell Mailchimp to Bending Spoons for $3.68B — $3.31B cash plus $368M in Bending Spoons stock — expected to close in early 2027. Intuit paid $12B for Mailchimp in 2021 and is taking the sale at less than a third of that; it says the divestiture sharpens focus on AI-native lead-to-cash tools for small businesses and will fund a buyback. Bending Spoons folds Mailchimp into a portfolio that already includes Evernote, Vimeo, WeTransfer, AOL and Airtable.
ChatGPT adds live dictation across apps, drops 10-minute cap
OpenAI is rolling out ChatGPT Live Dictation to all plans on iOS, Android, web and the desktop app, starting Nov 18. The feature transcribes speech in real time and eliminates the previous ten-minute recording cap, letting users dictate continuously instead of waiting for a batched transcript. Gizmodo flags the obvious downside: removing the cap plus always-on transcription could push users to volunteer much more personal data to the model.
National cyber director Cairncross steps down before Senate flip
Axios reports National Cyber Director Sean Cairncross is stepping down now so the White House has time to confirm a successor before Democrats take the Senate in January, following the November midterms that cost Republicans the chamber. Cairncross, who had led the Office of the National Cyber Director since confirmation, did not want to serve another two years; the departure is described as amicable. The ONCD coordinates federal cyber policy and has been the White House's point of contact on AI-system incident response and critical-infrastructure guidance.
BSI: 54% of bosses say AI is cutting headcount, up from 41%
British Standards Institute research published Nov 19, based on a survey of 1,287 business leaders across nine countries, finds 54% say AI is helping reduce headcount, up from 41% a year earlier, and 49% say AI has already reduced entry-level opportunities. Leaders also acknowledged using AI by default: 51% said they use AI without considering whether it is the best option, and 57% now reach for AI instead of consulting colleagues. BSI warns that short-term productivity gains may be coming at the cost of long-term talent pipelines.
Apple job listings reveal plans for a coding, browsing AI agent
Apple posted a Senior Engineering Manager role for a 'proactive, agentic AI assistant' that browses the web, works across apps, runs commands, writes and debugs code, and calls services while coordinating specialized sub-agents across apps, browsers, command lines, code and APIs. Bloomberg's Mark Gurman, who first flagged the postings, says Apple could introduce the capabilities as early as 2027 as it tries to catch Meta's Muse, ChatGPT and Claude Code. The language suggests early-stage development; Apple is now recruiting engineering leadership to run it.
House panel: Commerce OK'd 879,000 AI chip exports to China
House Foreign Affairs Chair Brian Mast on Nov 19 released Commerce export-license data showing BIS approved more than 879,000 Nvidia H200 and equivalent AMD AI chips for sale to China between December 2025 and May 2026, out of 3.4 million requested. Named recipients include Alibaba, Tencent, ByteDance, Tesla Shanghai and Baidu, plus two firms — Maginfra and New Matrix — the committee flagged as suspected smuggling fronts with empty offices. The disclosure is being billed as the most detailed public release in the history of the Export Control Reform Act.
Bajaj puts $4M into Soket AI for India's EKA-1 sovereign model
Bajaj Finance and Pennant Technologies each put $2M into Soket AI, backing a sovereign-AI model line anchored by EKA-1, a 26B-parameter foundation model focused on math, code and reasoning. The IndiaAI Mission-backed startup is now targeting larger models that cover 100+ languages, including Indian and Global South languages, and that run on-premises. Pennant plans to co-build banking-specific LLMs and agents for KYC, onboarding, collections, compliance and fraud detection.
Kokoro TTS HF account hijacked via lapsed domain, weights edited
Researcher Alon Hertz's team registered the lapsed hexgrad.com domain, used it to reset the Hugging Face account publishing Kokoro (a TTS model with ~13 million monthly downloads), and demonstrated they could fine-tune its weights in 87 seconds on a gaming GPU — turning, in one example, a cake recipe output into an omelette recipe. They did not re-upload the altered model, but noted that code review, static analysis, and signature scanning would not flag weight-only tampering. Kokoro is supported by tooling from Google, NVIDIA, Intel, AMD, and AWS.
Hilton and AOC press House to pass DEFIANCE Act on AI porn
Paris Hilton and Rep. Alexandria Ocasio-Cortez held a Capitol press conference pressing House Judiciary to advance the DEFIANCE Act, which would create civil liability for non-consensual AI-generated sexual imagery. The bill cleared the Senate unanimously in January, has 70 House cosponsors and backing from Microsoft, Snap and OpenAI. Hilton, chief advocacy officer at Loti AI, said 100,000+ deepfakes of her have surfaced online.
Kuaishou's Kling Image 4.0 generates native 8K from one prompt
Kuaishou unveiled Kling Image 4.0 on Nov 18 with native 8K image generation, up to 10 coordinated images per prompt, and 2,000 characters of in-image text across seven languages and 70+ commercial fonts. The model is available via API to developers and enterprise customers, targeting advertising, storyboards and e-commerce design, and follows Kling 4.0 video from October.
Snap ships CLI that builds AR lenses via Claude Code and Cursor
Snap launched Easy Lens CLI on Nov 18, letting developers pipe plain-language prompts through Claude Code, Cursor or Codex to build and publish Snapchat AR Lenses end-to-end. The tool runs on Mac, Windows and Linux and plugs into Snap's Lens+ Payouts revenue program, extending the company's bet that agentic coding assistants can turn a Lens idea into a shipping product in one day without hand-written code.
Ruby2Shell bug gives pre-auth RCE on stock Rails
Veria Labs published Ruby2Shell (GHSA-34m4-m9fh-gqp5), a one-byte heap overflow in Ruby's string case-conversion functions triggered by upcasing the German ß in legacy encodings; researchers demonstrated pre-authentication RCE on stock Rails in its default configuration. Affected: Ruby 3.3.0–3.3.13, 3.4.0–3.4.12, and 4.0.0–4.0.9. UTF-8 strings are safe, but upcase, swapcase, downcase(:fold) and casecmp? all reach the overflow.
DeepMind formally verifies Signal protocol with Gemini
Google DeepMind introduced AlphaProof Code, a multi-LLM agentic system in which Gemini Carbon agents generated over 6 million lines of Lean to formally verify Signal's protocol in four weeks versus years of manual work. Agents flagged only ~1% of the code for human review while Lean verified the rest, and the system identified two reintroduced bugs in isolated test builds. It is positioned as the first AI-driven formal verification of a widely deployed cryptographic protocol at implementation depth, with the verification artifacts open-sourced.
Students sue Omnilert after AI called Doritos bag a gun
Four Black Baltimore County students filed suit on Nov 18 in Baltimore County Circuit Court against Omnilert, BCPS, and the county police department after Omnilert Gun Detect flagged a Cool Ranch Doritos bag as a firearm at Kenwood High School in October 2025, prompting six squad cars and a gunpoint takedown. According to the complaint, officers on scene muttered 'stupid AI cameras' and 'fucking Omnilerts' while handcuffing the teens. Plaintiffs seek damages and a court order barring further use of the system until verification protocols are fixed.
Citi, PNC, Citizens: AI agent hack losses now the customer's problem
Citi, PNC and Citizens updated terms on November 18, 2026 to hold customers responsible for anything third-party AI agents do in their accounts — including activity after credentials are leaked or the agent itself is hacked. Citi carves out an exception for AI agents Citi provides directly. JPMorgan and Bank of America adopted similar policies last month, making five major US banks now pushing agent-hack liability onto end users.
GAO: ChatGPT, Claude, Gemini, Grok all planned terror attacks
The Government Accountability Office reports that ChatGPT, Claude, Gemini and Grok each produced viable US terrorist attack plans in under 12 hours after investigators bypassed each platform's safeguards. Two models produced IED plans aimed at mass casualties; two others produced plans targeting the electrical grid. The audit, requested by Sen. Maggie Hassan, ran during spring 2026 and deliberately withheld which model produced which plan to prevent roadmapping of bypasses. The incoming chair of the Senate Homeland Security Committee is preparing hearings on AI safeguards in response.
RTX Spark mini-PCs ship, pack 128GB unified memory for local AI
NVIDIA launched RTX Spark compact desktops on November 18 with ASUS, Dell, and MSI systems available immediately; Acer, GIGABYTE, HP, and a Microsoft Surface RTX Spark Dev Box follow. Each pairs a 20-core Grace CPU with a Blackwell GPU carrying up to 6,144 CUDA cores, 128GB unified memory, and ~1 petaflop of AI performance — enough to run 120B-parameter models locally. MSI's Edge Mesa N AI+ starts at $4,699.
TRM Labs lands $95M DHS deal to power new cybercrime center
The Homeland Security Task Force launched the Threat Disruption Center on November 18, naming TRM Labs as its primary private-sector partner under a reported ~$95M one-year no-bid deal. TRM will deploy its Ardence AI orchestration platform as the center's primary investigative tool, serving 30+ federal agencies and 440+ state and local partners tracing scam flows and sextortion networks. US cybercrime losses hit nearly $21B in 2025, with crypto-related theft making up more than half.
Google ships Gemini 4 Neon, ranks third on intelligence index
Google released Gemini 4 Neon, a frontier model priced at $3 per million input tokens and $15 per million output tokens, with a 1M token context and four configurable thinking levels. It ranks #3 on the Artificial Analysis Intelligence Index at 61, #2 on the Vals Index and #2 on Agent Arena for steerability, and is available via API in Google AI Studio and Android Studio, in Gemini for Business, in the Gemini app (Pro/Ultra), and in Google Antigravity.
Anthropic prices IPO at $83, above its $70-$78 range
Anthropic priced its IPO at $83 per share Wednesday, roughly 6% above the top of its previously announced $70-$78 range, with Nasdaq trading set to begin Thursday under the ticker ANT. The company upsized the offering amid strong institutional demand and is separately locking down $26B in commitments from Nvidia, Microsoft, Google, TSMC, Broadcom, AMD and Micron. The $83 pricing confirms the lift over earlier reports that Anthropic was weighing an upsize above $80.

This Week's Biggest Movers in AI

vs average of last 4 issues — click to explore

Entity Now Avg Change
Chips 115 3 ▲ +3733%
NVIDIA 104 0.8 ▲ +12900%
Agents 96 3.5 ▲ +2643%
Funding 62 0.5 ▲ +12300%
OpenAI 64 4.5 ▲ +1322%
Regulation 68 8 ▲ +750%
Anthropic 66 7.8 ▲ +746%
AI Infrastructure 58 2 ▲ +2800%
Jensen Huang 46 0.3 ▲ +15233%
Google 44 2.5 ▲ +1660%
Generative AI 37 0.3 ▲ +12233%
Inference 35 0.5 ▲ +6900%
Robotics 28 1 ▲ +2700%
Microsoft 29 3 ▲ +867%
Open Source 27 0.3 ▲ +8900%

How AI News Coverage Shifted This Week

News mix this week vs last

AI News Volume by Quarter

4-Issue Trend Lines: 113 AI Entities

Last 4 issues — click to explore