Friday, November 13, 2026

18 news stories · 24 video stories · 12 late finds

AI news

18 stories from written sources (3131 articles from 277 of 426 feeds, published between Nov 11 14:57 UTC and Nov 13 14:57 UTC, none shown in an earlier edition). Brief summaries with linked sources; the videos follow.

Top story

Nvidia says Jensen Huang has resigned as CEO; CFO Colette Kress is interim chief

Nvidia's board said Nov. 12 that Huang resigned as CEO, president and director effective immediately, days after he acknowledged an affair with a woman U.S. intelligence concluded was spying for China, the New York Times said. The Times reported a second relationship with an Nvidia executive. The board will consider internal and external successors. The Times said Nvidia lost more than $1 trillion in value this month; the FT said shares largely recovered by Thursday's close.

Sources: Axios · CNBC Technology · WIRED AI · Financial Times Technology · New York Times Technology

Labs and models

Google releases Gemini 4 Hydrogen, a lower-cost model it says scores 51 on an independent index

Google said Nov. 12 that Hydrogen replaces Gemini 3.8 Flash in the Gemini app, Search and its API, at $1.50 per million input tokens, according to The Next Web. On the Artificial Analysis Intelligence Index it scored 51 to Gemini 4 Argon's 50, per Google. The Decoder said it nearly matches Claude Sonnet 5.5 for less money but costs about 2.5 times as much per task as GPT-6.1 Sol.

Sources: The Next Web · SiliconANGLE · The Decoder · Ars Technica AI · Google DeepMind

Anthropic announces Claude Fable 5.2 and Mythos 5.2, with a system card

Mashable reported the announcement of the two models, The Batch listed them among Anthropic's new top models, and Zvi Mowshowitz reviewed the system card on Nov. 11. The available excerpts give no benchmark figures, pricing or availability.

Sources: AI Business · The Zvi · The Batch · Mashable

Policy and courts

Advocacy groups sue Anthropic, OpenAI, Google, Meta and a SpaceX unit in California to halt frontier model development

Four nonprofits filed suit Nov. 12 under California's Unfair Competition Law, MLex reported, arguing that the companies' pursuit of existentially risky AI is an unfair and unlawful business practice. CBS News reported the complaint alleges the developers cannot control their own technology. The excerpts do not give the court, the exact relief sought or the companies' responses.

Sources: Politico · MLex AI · CBS News · Reuters · AP News

Senate Banking chair asks Huang and Lutnick to testify Dec. 3 and Dec. 10 on Nvidia chip sales to China

Sen. Tim Scott sent letters Nov. 12 asking both to appear voluntarily, NBC News reported. Lutnick's spokesperson said the Commerce secretary agreed. TechCrunch, citing Bloomberg and NBC, reported House Foreign Affairs Chairman Brian Mast said Huang agreed to a closed-door appearance. The Financial Times reported Treasury Secretary Scott Bessent blamed Commerce for not blacklisting Megaspeed.

Sources: Financial Times Technology · TechCrunch AI · The Hill Technology · NBC News Tech · CNBC Technology

Bloomberg ties banned Nvidia chips to a Shanghai data center; Reuters finds more than $3 billion in Chinese projects built on restricted GPUs

Bloomberg reported, citing documents it reviewed, that a company founded by Alice Huang procured smuggled chips installed at a Shanghai data center linked to Megaspeed. Reuters said its review of Chinese contracts and disclosures found more than $3 billion in compute projects built or planned around restricted Nvidia GPUs. Nvidia has said export controls effectively shut it out of China's data-center market, Reuters noted.

Sources: Quartz · Engadget · Crypto Briefing · Reuters · Bloomberg Technology

EU leaders hold first dedicated AI summit in Dublin, with briefings from Mistral's Mensch and Hinton

Leaders met Friday, Nov. 13, for the first EU leaders' discussion devoted to AI, which Council President Antonio Costa called for, The Next Web reported. It said the agenda leads with competitiveness and funding and that a plan for AI gigafactories has stalled. Finland and Estonia urged EU-level arrangements for responding to serious AI incidents, Politico Europe reported.

Sources: The Hindu · DW.com · Politico Europe Technology · The Next Web · Reuters

Business and chips

Anthropic's IPO is about six times oversubscribed, Bloomberg reports; pricing set for Nov. 18

Bloomberg reported demand is about six times the shares on offer. In an amended filing Nov. 10, Anthropic set a range of $70 to $78 for 825 million shares; at the top that would raise $64.35 billion and value it at up to $1.405 trillion, CNBC reported. Anthropic plans to price Nov. 18 and begin trading Nov. 19. Quartz said a $26 billion private placement would add to the raise.

Sources: AI Business · Quartz · CNBC Technology · Bloomberg.com · Reuters

OpenAI raises $15 billion at a $1.25 trillion valuation, sources tell CNBC

CNBC reported the round has reached $15 billion, up from the $11 billion first tranche reported Nov. 7. Dealroom and 24/7 Wall St. describe it as half the original target; the earlier report had put the possible total at $20 billion, so the target figures differ. The Next Web reported quarterly revenue of about $10 billion.

Sources: 24/7 Wall St. · Benzinga · The Next Web · Dealroom · CNBC

Nscale's amended IPO filing shows $1.69 billion quarterly loss and a need for more financing

CNBC reported a third-quarter net loss of $1.69 billion on $307.5 million in revenue, up 1,344% from a year earlier. Bloomberg reported debt of $5.5 billion in September, up from $137.4 million in June, and closer ties to Nvidia. The Financial Times said Nscale told investors it needs more financing to fulfill Anthropic contracts.

Sources: WSJ · Quartz · Financial Times Technology · Bloomberg Technology · CNBC Technology

Research

Google DeepMind announces AlphaGenome 2, a model that predicts how DNA variants affect specific human cell types

DeepMind said Nov. 11 the model conditions predictions on tissue, cell type and cell state, extending last year's AlphaGenome. ITmedia reports it was trained on data from more than 100 million cells and is available through an API for noncommercial research, with a limited preview for commercial research.

Sources: Dataconomy · TechRepublic · ITmedia AI+ · Google DeepMind

DeepMind says AlphaProof Code verified the cryptographic core of Signal in four weeks

The Decoder reported that agents powered by Gemini Carbon wrote more than six million lines of proof code and flagged about 1% for human review, in work it said typically takes years. DeepMind plans to extend the method to infrastructure such as power grids and medical devices, according to The Decoder.

Sources: The SOFX Report · Dealroom · HPCwire · The Decoder

Tools and open source

Anthropic launches Claude Meetings beta, adding Claude to Google Meet calls for Team and Enterprise plans

Anthropic began the rollout Nov. 12, Unite.AI reported. Claude appears as a visible participant, answers in writing by default, uses connected tools with permission and builds notes, documents and whiteboards during the call. The Decoder said more platforms are to follow.

Sources: Dataconomy · ZDNET AI · The Decoder · Unite.AI

OpenAI launches Daybreak Red Team and Daybreak Defense Factory security tools

Announced Nov. 12 at OpenAI's Cyber at the Frontier event, per SecurityWeek. Red Team uses agents to test web applications and confirm vulnerabilities; Defense Factory traces findings and drafts patches that engineers review before merge. India Today and The Register reported OpenAI also warned organizations that skipping such tools risks being hacked.

Sources: India Today · The Register · Help Net Security · SecurityWeek

Safety and security

Unit 42 says Chinese-speaking attackers used the Cairn AI framework in 25 operations against at least 19 organizations

Palo Alto Networks' Unit 42 said the attackers paired the open-source penetration-testing framework with Zhipu GLM and DeepSeek models in August and September, supplying a target and a goal. It said access or data retrieval succeeded at 10 organizations and assessed with moderate confidence that the activity is China-nexus intelligence collection.

Sources: GBHackers News · Help Net Security · CSO Online · Unit 42 (Palo Alto)

Notion says Fireworks AI breach may have exposed Notion AI chat records, including for some zero-retention customers

Notion Labs announced Nov. 9 that unauthorized access at vendor Fireworks.ai may have exposed chat records from its Insights feature, according to ITmedia. Notion later said enterprise customers on zero data retention could be affected because the Fireworks endpoint it used had retained data. TechCrunch reported the intrusion Nov. 12.

Sources: Cryptonews.net · ITmedia AI+ · TechCrunch

VulnCheck says Project Glasswing has produced few fixes after seven months

Citing Anthropic's disclosure dashboard dated Nov. 2, VulnCheck's Patrick Garrity said researchers found 37,966 vulnerabilities, 7,349 of them in Anthropic's public ledger. Infosecurity Magazine said only 217 have been fixed; The Stack reported two were exploited. Anthropic's response is not stated.

Sources: The Stack · Infosecurity Magazine

Microsoft says AI-assisted MDASH project yielded 59 Linux kernel bug reports, 33 CVEs and 38 merged fixes

Microsoft's FORGE Lab said scanning produced 26,502 potential findings in about two weeks, narrowed by screening and human validation to 59 upstream reports, and demonstrated eight local root privilege-escalation exploits. CSO Online reported some bugs had survived more than 20 years and that validation and patching stayed largely human work.

Sources: CSO Online · Microsoft Security Blog

Get the next edition by email, or read on for the videos.

Then, from YouTube

What AI YouTube said

Coverage: 481 videos from 88 of 111 monitored channels reviewed (0 partial or status unknown); videos under 45 s were not reviewed. Dates are UTC upload dates. Vendor claims are labelled as such.

Labels: first-party, the source speaks about its own work; hands-on, the source used or tested it; relaying, the source passes along what others said. They are not accuracy ratings.

New today

Anthropic releases Claude Fable 5.2; reviewers report higher per-task cost and faster limit use than Opus 5.5

Anthropic released Claude Fable 5.2 on Nov. 9, 2026, and said Mythos 5.2, offered to vetted organizations, is the same underlying model with different classifiers and access checks. An Anthropic employee, speaking at AI Engineer, called Opus 5.5 her preferred daily driver and Fable 5.2 strong at verifying its own output; Julian Goldie, relaying Anthropic, reported 69.9% on Frontiers SWE versus 62.3% for Opus 5.5, and The AI Advantage said Anthropic reported Fable completing 41% of Remote Labor Index projects versus 21% for Opus 5.5. Nate B Jones said the API price is $8 per million input tokens and $40 per million output tokens, a figure he stated without showing a price page. Reviewers reported higher spend: Cole Medin's presenter cited Artificial Analysis cost per task of $26.54 at max and about $9 at high versus $3.46 for Opus 5.5, David Ondrej said he hit usage limits 53 times in three days on Fable versus twice on Opus across four $200 plans, and Theo said Fable is capped near half of weekly usage on the $200 plan. In Dylan Davis's test, Opus 5.5 and Fable 5.2 at medium effort found the same nine planted issues in 15 files.

Evidence: 2 first-party, 7 hands-on, 1 relaying

Disagreements: Per-task cost comparisons differ by presenter and setting: Dylan Davis claims at least twice Opus 5.5 per task, up to 3-6x; Theo reads guidance that Fable 5.2 at high beat Opus 5.5 at max at half the per-task cost; the Artificial Analysis figures cited by Cole Medin's presenter put Fable at high near $9 versus $3.46 for Opus 5.5 without a stated Opus setting. The conditions are not reconciled in the items.

Watch: Dylan Davis: Claude Fable 5.2: Anthropic Says Use Opus for Most Work; Theo - t3.gg: Stop Picking The Wrong Model

Channels relay reports that Nvidia CEO Jensen Huang resigned; spy and export allegations unproven

Three channels said on Nov. 12, 2026 that Nvidia announced Jensen Huang's resignation as CEO with immediate effect and named its CFO interim CEO; Matt Wolfe said Nvidia announced it, while David Shapiro and Fireship relayed it secondhand. No Nvidia statement is shown in any video, and the items do not verify it. Fireship and Wolfe also recapped press reports of a probe into a Singapore buyer, called Mega Speed in captions, over more than $4.8 billion of Nvidia hardware, plus a short seller's report; Wolfe said the spy allegations are unproven, that China calls them a smear and that Nvidia's board says it has no reason to believe secrets leaked. The Mastra hosts, recording earlier that day, discussed a leaked letter and an alleged spy link per social posts and said they know nothing beyond what was posted. Shapiro's market view is labelled off-the-cuff and not financial advice.

Evidence: 0 first-party, 0 hands-on, 4 relaying

Disagreements: Details differ across relays: Fireship says the board statement came five days before the earnings call and that stock fell another 10%, Wolfe says stock was down over 13% at one point, and the Mastra recording predates the resignation reports.

Watch: Matt Wolfe: MASSIVE AI Shake-Up That Changes Everything

Google releases Gemini 4 Hydrogen, a Flash-tier model priced at $1.50 per million input tokens

Google released Gemini 4 Hydrogen, the fast tier of the Gemini 4 family, in AI Studio and the Gemini API, reviewers said; Prompt Engineering and Bijan Bowen put the price at $1.50 per million input tokens, and Bowen at $7.50 per million output tokens versus $2 and $10 for Sonnet 5.5. Prompt Engineering's presenter said it scores one point above Gemini 4 Argon on the Artificial Analysis index, and Fahd Mirza said Google's table has it ahead of Sonnet 5.5 in most rows but behind in three. In hands-on tests Mirza had it find a currency bug in about three to four minutes, and Bowen said a robot-arm test worked at extra high effort but took about an hour, with all tests using 4% of weekly usage on the $200 Ultra plan. Mirza's audio caption garbled the price, so the $1.50 figure rests on the other two sources.

Evidence: 0 first-party, 2 hands-on, 1 relaying

Watch: Bijan Bowen: Gemini 4 Hydrogen Is the BEST Flash Model Yet – Full Test!; Fahd Mirza: Gemini 4 Hydrogen: Everyday Model for Everyone from Google

Anthropic adds Claude Whiteboards, Sheets, Meetings and generally available Data on Nov. 10

Anthropic released Claude Whiteboards, Claude Sheets and Claude Data on Nov. 10, 2026, according to Julian Goldie: whiteboards and Sheets are in beta, Data (formerly Dashboards) is generally available, for Pro plans and up plus Team and Enterprise. Bart Slodyczka said whiteboards are in beta on all paid plans, count toward normal usage, and export to PNG or go to Miro, Lucid or Zoho. Anthropic introduced Claude Meetings, in which Claude joins video calls and builds artifacts live; Goldie said it is rolling out in beta for team plans, starting with Google Meet. Goldie reported Sheets has a 100,000-cell limit and no pivot tables, VBA or macros; that figure is unconfirmed by Anthropic in the items.

Evidence: 1 first-party, 1 hands-on, 1 relaying

Disagreements: Plan availability for whiteboards is stated as all paid plans by Slodyczka and as Pro and up plus Team and Enterprise by Goldie; the two are not necessarily inconsistent.

Watch: Bart Slodyczka: Claude's New WHITEBOARDS Are Amazing (Master it in 16 minutes); Julian Goldie: This New Claude Update Is Absolutely WILD! 🤯 (high hype)

Continuing stories

Google opens Gemini 4 Carbon to developers and Google Cloud on Nov. 10, adds Antigravity /swarm preview

Google opened Gemini 4 Carbon on Nov. 10, 2026 in AI Studio, the Gemini API and Google Cloud, according to Julian Goldie, who said Google first showed it on Oct. 28 to a small group of cyber defenders; the Gemini app gets it only for AI Ultra subscribers, still rolling out.

Evidence: 0 first-party, 1 hands-on, 1 relaying

Watch: Julian Goldie: Google Antigravity Just Got Swarm Mode (9 AI Agents, 1 Prompt) (high hype)

OpenAI reports GPT6 Cyber scored 52% on Exploit Gym and opens Daybreak Red Team early access

OpenAI said GPT6 Cyber scored 52% on its Exploit Gym benchmark and previewed a next cyber model it described as nearly double Mythos 52 on that benchmark.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: OpenAI: Cyber at the Frontier

Manolo Remiddi cites $2.9 trillion one-day S&P 500 loss and Nvidia down 24% in a month

Manolo Remiddi said the S&P 500 lost $2.9 trillion in one day, the largest dollar loss on record, and that Nvidia fell 24% in a month from a $239.24 peak; the figures are unsourced on screen and his dates are internally inconsistent.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Manolo Remiddi: AI Works. So Why Is the Market Panicking?

Anthropic speaker cites auto mode study and describes Claude Code hooks, CLAUDE.md and projects beta

At AI Engineer, Anthropic's Hallie cited an August study in which humans caught about 14% of dangerous commands while auto mode blocked 89%, and said auto mode became Claude Code's default permission mode in August.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: AI Engineer: Rethinking Claude Code Best Practices — Lydia Hallie, Anthropic

Alibaba announces Qwen 4 family and Qwen 4 Max preview without size or benchmarks

Alibaba's Qwen team announced the Qwen 4 family on Nov. 11, 2026 and released a Qwen 4 Max preview, Julian Goldie said; as of that date Alibaba had not disclosed its size, benchmarks, release date or weight openness.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Julian Goldie: Alibaba's NEW Qwen4 Is Coming! (high hype)

Julian Goldie recaps Claude Dashboards beta, Motion, and Docs, Slides and Design leaving beta

Julian Goldie recapped Anthropic's Claude Dashboards beta, which he dated Oct. 8, along with Claude Motion and Docs, Slides and Design leaving beta; the items add no new information beyond earlier coverage.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Julian Goldie: Claude's New Dashboards Are ABSURD! (high hype)

Also notable

Google releases Gemini Decision 1, a model that returns scored answers instead of text, with a free AI Studio tier

Google released Gemini Decision 1 and a Decisions API, with a free tier in AI Studio, according to Julian Goldie; the model scores yes/no, pick-one or scale questions in one pass with a confidence value and no output text.

Evidence: 0 first-party, 2 hands-on, 0 relaying

Watch: Fahd Mirza: Google’s Gemini Decision 1: Fast, Cheap, Not Perfect

Perplexity reports no VM escapes in nine-model sandbox tests; four models bypassed a network gateway

Perplexity, as relayed by Julian Goldie, reported no virtual-machine escapes in part one of a study of nine models across 216 runs, 108 per condition, in its open-sourced Firecracker-based sandbox.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Julian Goldie: Perplexity Tried to Make 12 AI Models ESCAPE?! 🤯 (high hype)

Meta launches Muse Amber, its smallest Muse model, with weights kept private

Meta launched Muse Amber on Nov. 9, 2026, announced by Mark Zuckerberg on X, and is keeping its weights private, according to Julian Goldie.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Julian Goldie: Meta Just Released a CRAZY CHEAP AI Model?! (high hype)

OpenAI launches businesses in ChatGPT in beta, including sponsored business agents for labeled ads

OpenAI launched businesses in ChatGPT in beta on Nov. 10, 2026, with a Business Hub and insights, according to Julian Goldie.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Julian Goldie: OpenAI Just Changed ChatGPT SEO FOREVER?! 🤯 (high hype)

Google DeepMind announces Gemini Robotics 2.5, reporting 96% success in a Mercedes-Benz kitting pilot

Google DeepMind announced Gemini Robotics 2.5, a vision-language-action model, on Nov. 10, 2026, per Julian Goldie, who said it is rolling out to robotics partners with no public access details.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Julian Goldie: Google's AI Robot Just Hit 96% Accuracy?! (high hype)

GitHub makes stacked pull requests generally available on github.com, with auto-merge and worktree support coming or added

GitHub said stacked pull requests are generally available on every github.com plan, with Enterprise Server later, and reported that repositories using them see 9% more merged code than peers, a GitHub-reported figure without stated method.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: GitHub: Stop waiting on your own pull requests: GitHub stacked PRs in practice

Microsoft Research roundup reports KLS conjecture resolution and a 13.5 million-session compaction cost study

A Microsoft Research host said new work resolves the 31-year-old KLS conjecture, that results were independently validated and the proof formally verified with computer-assisted methods.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: Microsoft Research: Microsoft Research Focus – Edition #98

Two Minute Papers reports AlphaProof Code produced 6.5 million lines of Lean proofs for libsignal

Two Minute Papers reported AlphaProof Code produced 6.5 million lines of Lean proofs for libsignal in four weeks and caught both re-introduced known bugs in a stress test.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Two Minute Papers: DeepMind’s AI Just Did The Impossible

Models and learning

Anthropic engineer describes thinking protections, mid-conversation effort changes and prompt caching guidance

In a Claude video, Anthropic engineers said the company documented new thinking protections as a measure against distillation, that the API allows effort and system prompts to change mid-conversation without breaking prompt caching, and that Anthropic is moving from extended to adaptive thinking with a temporary toggle back.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: Claude: Building the Claude API for the Frontier

Databricks demo compares Lakebase with Amazon RDS on provisioning, branching and migration times

In a Databricks demo, a Lakebase database was ready in 13 seconds versus 9 minutes 39 seconds for RDS, and a branch of a 1-million-order database took 8 seconds versus 12 minutes for an RDS restore.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: Databricks: Compare Amazon RDS for Postgres with Databricks Lakebase

Sentry says MCP carries about 99% of headless usage and its Warden harness has cost about $700,000

On a Mastra livestream, Sentry said MCP carries about 99% of its headless usage, far above its CLI, and that its custom security code-review harness Warden has cost about $700,000 so far, with seven figures projected next year.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Mastra: Jensen’s Apology, Fable 5.2 & MCP vs CLI

Andrew Ng says experienced engineers steer coding agents best and AI excels at verifiable tasks

In a Neo4j-hosted talk, Andrew Ng said vibe coding without fundamentals yields poorly chosen trade-offs, that AI excels at verifiable tasks while judgment tasks lack RL gyms, and that fully automated ontology construction still underperforms human-steered approaches.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Neo4j: Live from NODES 2026 - Fireside Chat with Andrew Ng and Emil Eifrém…

Vizuara reviews DreamZero paper claiming 38x inference speedup and tests SO-101 fine-tuning cost

Vizuara said the DreamZero paper claims a 38x inference speedup enabling real-time control at seven frames per second and cross-embodiment transfer from minutes of unpaired video, which the speaker did not test.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Vizuara: Lecture 15 - World Action Models | DreamZero from Scratch (+ SO-101…

Imagination Studio is a free open-source local AI film studio with non-commercial model licenses

Julian Goldie said Imagination Studio is a free open-source local AI film studio for Windows, Apple Silicon Mac and NVIDIA Linux, whose Mac installer reportedly needs 48 to 128 GB of memory depending on features.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Julian Goldie: Imagination Studio: NEW FREE AI Film Studio Is INSANE! 🤯 (high hype)

Late finds

377 videos uploaded Nov. 9 to early Nov. 12 (before 03:14 UTC on Nov. 12), earlier than this edition's 24-hour window. The editions for Nov. 10, 11 and 12 were not produced, so these transcripts were read only now. Stories already covered above (Claude Fable 5.2, Gemini 4 Carbon, Muse Amber, Claude Sheets and Whiteboards, the Nvidia reports) are not repeated. Selected stories drawn only from these videos:

Fireship relays OpenAI report of an agent run producing 948 unique Linux root exploits at about $1,280 each

Fireship, relaying OpenAI's publication, said agents working on the Linux master branch across about 5,000 audit targets confirmed 1,724 bugs and produced 948 unique root exploits with patches, at a total cost of $1,213,000 or about $1,280 per exploit.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Fireship: OpenAI spent $1.2 million hacking Linux... and it was a bargain

System card: UK AISI says Mythos 5.2 flagged possible alignment testing in 96% of trajectories

AI Explained, citing the 264-page Fable and Mythos 5.2 system card, said the UK AI Security Institute found Mythos 5.2 considered or identified in its chain of thought that it was in an alignment evaluation in 96% of trajectories, higher than other Anthropic models.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: AI Explained: Fable 5.2: The Self-Improvement Loop Keeps Loopin’ … (plus Gemini…

Anthropic says it limited Fable 5.2 gains in AI-research skills; METR report discussed

Anthropic said it took steps in training to limit Fable 5.2 and Mythos 5.2 advances in research capabilities that could most directly accelerate AI development, which Wes Roth read as keeping accelerator-kernel optimization and LLM-architecture improvement near Opus 5.5 level.

Evidence: 0 first-party, 0 hands-on, 3 relaying

Watch: AI Explained: Fable 5.2: The Self-Improvement Loop Keeps Loopin’ … (plus Gemini…

Zenity reports about 100 malicious agent skills and a fake 'Paperclip' campaign tied to 1.7 million downloads

Zenity said in an AI Engineer talk that a lookalike GitHub organization cloned the Paperclip project and pushed a payload collecting SSH, AWS and Azure credentials, with 1.7 million total downloads by the speaker's figure, and that everything was down within 12 hours of reports to Vercel and Microsoft.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: AI Engineer: You Can't Tell What an Agent Skill Will Do by Reading It - Roey Ben…

NVIDIA releases Nemotron 3.5 Bolt, a 74B-parameter MoE with 2-bit experts that fits in 24 GB

Sam Witteveen said NVIDIA's Nemotron 3.5 Bolt 74B A9B, pruned and distilled from Nemotron 3.5 Super VL, has about 9B active parameters and 2-bit experts, shrinking from 148 GB to 22 GB.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Sam Witteveen: NVIDIA's Nemotron Bolt 74B-A9B 2 Bit

Google announces two eighth-generation TPU inference systems, including a 384 GB HBM design

Google's Amin Vahdat announced a confidential-inference TPU system built on the 8t silicon base with 12.6 petaflops native FP4 and 216 GB HBM, and a high-memory system with 384 GB HBM, 17.2 TB/s bandwidth and 768 MB SRAM.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: Google Cloud Tech: The Future of AI Infrastructure

After midterms, Democratic lawmakers call to ban recursive self-improvement or pause advanced AI

David Shapiro said Rep. Greg Casar, who leads the Progressive Caucus, posted "ban RSI" three days after the midterms, and Sentdex said Rep. Yasamin Ansari tweeted that pausing advanced AI development should be the top Democratic agenda item; both are secondhand accounts of posts.

Evidence: 0 first-party, 0 hands-on, 2 relaying

Watch: Sentdex: AI Bubble over? - Happenings in AI

Cognition releases SWE 2 Flash, a 276B-total, 12B-active coding model, and withholds weights

Cognition released SWE 2 Flash on Nov. 4, 2026 in Devin desktop and Devin CLI, free until Nov. 30, according to Julian Goldie relaying the launch.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Julian Goldie: Cognition's NEW SWE-2 Flash Is INSANE! 🤯

Binance says attackers used internal AI tool after fake job-test malware download

Binance said an engineer who believed he was interviewing for an outside job downloaded malicious code to a work computer as a technical test, and that attackers then used an internal AI tool to reach two wallet engineers' computers in the Oct. 29 theft, according to Nate B Jones.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Nate B Jones: China Won't Take 600,000 Nvidia Chips. Did It Target Jensen Huang For…

OpenAI offers its Codex harness as a managed Agents API and announces computer use over the API

OpenAI presenters said the Agents API is a managed service built on the Codex harness, letting developers define tools and instructions and choose an OpenAI-hosted, own or third-party sandbox, and that computer use over the API with Astra was introduced the same day.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: OpenAI: Harness-as-a-Service: Building with the Agents API | DevDay 2026

OpenAI relaunches Codex Security Cloud on Daybreak Blue and details gated access for cyber defenders

OpenAI said Codex Security Cloud was rebuilt on the Daybreak Blue model, with no access request required for the cloud product, and that the Codex Security CLI and SDK are open source under Apache 2.0 on NPM.

Evidence: 1 first-party, 0 hands-on, 1 relaying

Watch: OpenAI: How to request and enable Daybreak Blue

OpenAI presents GPT Live 1, a full-duplex voice model, and an API with delegation to backends

OpenAI presenters described GPT Live 1 as a voice model that streams audio in and out at once and handles turn-taking without a separate voice-activity detector, with promptable tone and eagerness.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: OpenAI: Building with GPT-Live and Image Generation | DevDay 2026

Coverage

481 videos with transcripts reviewed (all fully read), 556 inventoried in the Nov. 9 to Nov. 13 scan window; 104 fall in the 24 hours ending Nov. 13, 03:14 UTC, and 377 are late finds. Not reviewed: 52 under 45 seconds, 15 live or upcoming streams deferred, and 18 without a saved transcript. 59 new batches; 5 videos were already extracted in earlier editions. The collector stopped on a YouTube access block (RequestBlocked) after saving 104 of 122 queued transcripts; the remaining 14 were not attempted and were not retried. News: 8 of 426 feeds failed and 16 were skipped as dead.

Sources reviewed (481 videos)

All 481 videos read for this edition, grouped by channel. Tags: what the video was, plus sponsored (paid placement), vendor (a company on its own product), self-promo (the creator's own offer), high hype or partial transcript where they apply.