Monday, October 12, 2026

18 news stories · 21 video stories

AI news

18 stories from written sources (1762 articles from 170 of 438 feeds, published between Oct 10 09:01 UTC and Oct 12 09:01 UTC, none shown in an earlier edition). Brief summaries with linked sources; the videos follow.

Top story

Trump administration reverses request that AI developers withhold unreleased models from UK safety institute, Telegraph reports

The administration dropped an earlier request that developers such as OpenAI and Anthropic not share new models with Britain's AI Security Institute before release, the Telegraph reported Sunday, Oct. 11, according to Reuters. Spokespersons for the White House, OpenAI and Anthropic did not immediately respond to Reuters. The reversal follows a Friday statement from a new White House AI task force that developers must report security incidents.

Sources: Bradenton Herald · Crypto Briefing · Firstpost · LBC · Economic Times Tech

Policy and courts

D.C. Circuit upholds Anthropic supply-chain-risk designation under FASCSA, National Law Review reports

According to The National Law Review, the D.C. Circuit upheld the federal supply-chain-risk designation of Anthropic and endorsed a broad reading of the statute. The shortlisted excerpt did not give the panel, the date or the reasoning.

Sources: The National Law Review

UK AI minister to press frontier labs on safety as ministers explore compulsory reporting, FT and Guardian report

Kanishka Narayan will say Britain cannot rely entirely on technology companies to police their own models and will announce stronger cyber defences, the Guardian reported. The Financial Times headlined that ministers are exploring compulsory safety reporting and said Narayan will tell companies to exercise stricter control over autonomous models. Neither excerpt gave a bill or a timetable.

Sources: Financial Times · The Guardian AI

Microsoft's Brad Smith says AI is less regulated than milk and calls for 'kill switches'

Microsoft President Brad Smith told the BBC that a washing machine faces more safety rules than AI and that companies need emergency brakes to pause or slow systems that grow out of control. He also urged a legal obligation to victims of rogue agents, the Guardian reported. His remarks follow Satya Nadella's weekend call for a human-controlled emergency brake.

Sources: Cybernews · BBC · The Guardian AI

Financial Stability Board chair sets out work on AI and stablecoins in letter to G20 finance chiefs

The board published its chair's October letter to G20 finance ministers and central bank governors on Oct. 11. Its title says the letter sets out the board's work on AI and stablecoins, and TradingView described it as flagging risks from both. The excerpts gave no details of the AI findings.

Sources: TradingView · Financial Stability Board

Business and chips

Amazon in talks to buy Israeli-founded AI startup Decart for about $7 billion, WSJ reports

Amazon is negotiating to acquire Decart for around $7 billion, and a deal could come within days unless talks collapse or another bidder emerges, the Wall Street Journal reported, per SiliconANGLE. Decart makes software for running AI models on different chips. Techmeme's summary of the report says Anthropic passed on buying it for about $6 billion in September.

Sources: CTech · The Economic Times · Bellingham Herald · SiliconANGLE · Bloomberg Technology

Intel and Broadcom chief executives say capital may become AI's biggest constraint

Broadcom CEO Hock Tan said financing could become a bigger obstacle than chips, memory or electricity, speaking Monday at the DBS Asia Future Forward Conference in Singapore, CNBC reported. Intel CEO Lip-Bu Tan pointed to the cost of capital and investors' demand for returns. CNBC noted that higher bond yields could raise the cost of funding data centers.

Sources: Sacramento Bee · CNBC Technology

IMF says AI investment keeps global growth near 3% and its chief warns of jobs and tax disruption

The IMF said AI investment is holding global growth around 3%, with major benefits for Asia, Reuters reported, and that productivity gains could help offset Asia's ageing population. Bloomberg reported the IMF's head called AI a "slow-moving tsunami" for jobs and taxes and urged policymakers to rethink education.

Sources: Bloomberg Technology · Reuters

Residents contest Google's $15 billion AI data center in India's Visakhapatnam

AFP reported that excavators are clearing land for Google's largest data center outside the United States, and residents fear strains on water and power and too few jobs. Google's Alexander Smith called the project a multi-decade commitment with "an obligation to be a good neighbour." The Andhra Pradesh government told AFP that displaced residents get compensation and new housing.

Sources: Tech Xplore · Economic Times Tech

Senior Iranian official threatens regional data centers and AI facilities if war escalates, AP reports

The Associated Press reported, via Fortune, that a senior Iranian security official warned Tehran could attack cables, data centers and AI facilities in the region if fighting intensifies sharply again. He called the latest U.S. response to a proposal to end the war "insufficient."

Sources: Fortune

Research

Northwestern study finds AI-written security patches often pass crash tests but fail maintainers' checks

Researchers tested AI coding agents on 112 real security bugs from 84 open-source projects, Help Net Security reported. With Claude Opus 4.7, a third of patches that stopped the crash still failed the tests maintainers wrote for the real fix. The authors say benchmarks that only check the crash can overstate success.

Sources: Help Net Security

Paper led by ByteDance researchers says DeepSeek's memory-saving method can hurt retrieval accuracy

The study, which has not been peer-reviewed, found retrieval accuracy of DeepSeek-V4 models varied by up to 40 percentage points with the position of information in roughly 128,000-token inputs, the South China Morning Post reported. The authors tie the swings to how the models compress memory and note their tests were controlled and used limited coding examples.

Sources: South China Morning Post Tech

Tools and open source

OpenAI to add invisible watermark to ChatGPT text in the EU, Malay Mail reports

Malay Mail reported the watermark will be applied in the EU and not by default in Malaysia. Pasquale Pillitteri, a blog, said OpenAI describes it as textGrain with an API option and reported OpenAI detection rates of near 68% for English and 42% for Romanian. OpenAI's own announcement was not among the items.

Sources: Pasquale Pillitteri · Malay Mail

Mistral unveils ML4 open-weight model and says it holds an edge over Chinese rivals

According to Digitimes, Mistral introduced ML4, an open-weight model, and touts an advantage over Chinese competitors. The claim is Mistral's; the excerpt gave no benchmarks, size or license terms.

Sources: digitimes

Anthropic halves Sonnet 5.5 cache-read price, MIXED reports

According to MIXED Reality News, Anthropic cut the price of cache reads for Sonnet 5.5 by half, and its own pricing tool showed the old price for two days afterward. Anthropic's announcement was not among the items.

Sources: MIXED Reality News

Safety and security

CyberXero campaign used Claude Code to steal more than 628,000 Ukrainian records, researchers say

According to The Hacker News, researchers described a two-pronged, AI-augmented intrusion campaign against WordPress sites, e-commerce platforms and critical-infrastructure entities in Ukraine. It paired commodity tools such as Cobalt Strike with an AI orchestration layer built on Anthropic's Claude Code.

Sources: The Hacker News

Anthropic's OSS Scanner has flagged about 29,000 possible bugs, 516 fixed, New Stack reports

The New Stack reported that the scanner, launched last week, sends unreviewed Claude vulnerability reports to open-source maintainers and that finding bugs is proving far easier than fixing them. Per the outlet's headline, only 516 of roughly 29,000 possible bugs have been fixed.

Sources: The New Stack

Korea's Financial Security Institute says AI agent scanned iM Bank servers before web-shell attack

According to data the institute gave a lawmaker, as reported by Seoul Economic Daily, it detected an AI agent-based attack on iM Bank on Aug. 28. The attacker exploited an Oracle WebLogic flaw to upload a web shell and tried to map internal servers. The outlet said no data was leaked.

Sources: Seoul Economic Daily Technology

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

Then, from YouTube

What AI YouTube said

Coverage: 75 videos from 21 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

Google presents Gemini agent for work at Gemini at Work event on Oct. 8

Google Cloud announced a Gemini agent for business, described as Gemini Enterprise becoming a universal agent for work, at its Gemini at Work 2026 event on Oct. 8, according to three channels that relayed the keynote. Google said the agent runs in the cloud around the clock, continues jobs for hours or days after a laptop is closed, keeps four memory types, connects through MCP to tools such as Slack, Jira and Salesforce, and can send tasks to Gemini or Anthropic Claude models. Google claimed in its own testing that smart routing comes close to running everything on a frontier model at a third of the cost and 40% faster, and described per-agent cryptographic identity, an audit trail, an agent gateway firewall and co-worker agents with their own email, calendar and drive. In a Google demo, an agent-built churn-style model reported 90% accuracy and 85% precision. The keynote also referred to a new frontier Gemini model announced the week before; presenter Julian Goldie said he could not confirm the claim that the agent works non-stop.

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

Watch: Liam Ottley: Reacting to Google's "Gemini at Work" Launch: BEST AI PLATFORM!!; Julian Goldie: NEW Gemini Agent is Absolutely WILD! (high hype)

Continuing stories

OpenAI rolling out GPT-6 in ChatGPT with Soul for paid plans and Luna for Free and Go

Julian Goldie said OpenAI began rolling out GPT-6 in ChatGPT on Oct. 7 and continued Oct. 8, with Plus, Pro, Business and Enterprise users getting GPT-6 Soul and Free and Go users GPT-6 Luna, and Enterprise access dependent on admin settings.

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

Watch: Julian Goldie: NEW ChatGPT Update is CRAZY! (high hype)

StepFun Step 5 Preview offered free for a week, with vendor-reported coding scores

StepFun's Step 5 Preview, a 600B-parameter mixture-of-experts coding model with 27B active parameters, was released Oct. 8 and is on OpenRouter, with a week of free access in several coding tools, according to Julian Goldie.

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

Watch: Julian Goldie: NEW Step 5 Preview is CRAZY GOOD! 🤯 (high hype)

Panel recaps OpenAI math manuscript release and mathematicians objection

Panelists on David Shapiro's show and David Ondrej repeated the earlier-reported release of OpenAI's math manuscripts, giving counts of roughly 370, over 700 and 722, and relaying that almost 900 mathematicians signed a letter objecting, plus a Navier-Stokes claim of 10,000 agents over 88 hours.

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

Watch: David Shapiro: Dots, Bots, and Math! Oh my!

Snyk demos AI BOM and Evo inventory policy for models, skills and MCP servers

Snyk speakers demonstrated an AI bill of materials that listed 69 assets, including a Llama 3 8B with one critical and two high vulnerabilities, and Evo, which flagged a Salesforce model under a policy.

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

Watch: AI Engineer: AI Security Engineer Foundations + Certificate — Javier Garza, Snyk

Ondrej claims $200 Claude Code plan gives 16 to 20x the usage of Codex plan

David Ondrej claimed a $200 Claude Code plan gives 16 to 20x the usage of a $200 Codex plan, citing a SemiAnalysis figure of about 10.6x, and that Anthropic subscriptions are subsidized about 90%; the estimates are unsupported and extend Theo's earlier plan-value estimates.

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

Watch: David Ondrej: Agentic Engineering in 2027 will look completely different (high hype)

Panel: Grokbot gets its own email address; panelist says it planned a Scotland trip

Panelists on David Shapiro's show said Grokbot received its own email address the day before recording and that one panelist's two-week Scotland trip request produced a booking plan that stopped before payment; the host was disappointed by ChatGPT Dot.

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

Watch: David Shapiro: Dots, Bots, and Math! Oh my!

Also notable

Perplexity releases pplx-embed-v2-late late-interaction embedding models under MIT license

Perplexity released pplx-embed-v2-late on Oct. 7, 2026 in 0.6B and 9B sizes, MIT licensed on Hugging Face, according to Julian Goldie.

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

Watch: Julian Goldie: Perplexity Just Dropped Its New Embedding Models (high hype)

Google DeepMind opens SynthID detector to everyone worldwide in English on Oct. 7

Google DeepMind opened its SynthID detector to the public worldwide, in English, on Oct. 7, 2026, according to Julian Goldie; it was previously limited to media professionals and testers.

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

Watch: Julian Goldie: Google Just Made AI Detection 10X Better! (high hype)

Microsoft Decision One, a non-local fast decision model, reports 83.5% average accuracy

Microsoft released Decision One, a non-local model for fast structured decisions such as routing, classification and agent guardrails; Fahd Mirza relayed Microsoft's charts showing 83.5% average accuracy across 147,000 questions, against 82.3 for Jev, with a median latency of 85 ms.

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

Watch: Fahd Mirza: Microsoft Decision-1: Model for Fast Decision-Making

Google Research releases TabFM zero-shot tabular model with non-commercial weights

Google Research released TabFM, a model that predicts unlabeled rows from labeled ones in one forward pass without training on the user's data, with non-commercial weights, according to Prompt Engineering.

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

Watch: Prompt Engineering: Google's GPT-3 Moment for Databases & Spreadsheets (TabFM)

Ravenwolf reports Terminal-Bench 2.0 runs: GPT 5.5 with Codex first, Fable refused 13 tasks

Benchmarker Ravenwolf said in his Terminal-Bench 2.0 runs (Terminus harness, five runs per configuration) that GPT 5.5 with Codex ranked first and Gemini 3.5 Flash came second, close to GPT 5.5 but using 1.65 billion tokens against 136 million and costing $620 against $497 at published prices.

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

Watch: AI Engineer: From Zero to Leaderboard: Agent Evaluation — Wolfram Ravenwolf…

Panelist says Trump signed AI industry self-regulation framework with about 20 leaders

A panelist on David Shapiro's show said Trump signed what he called a super intelligence constitution surrounded by about 20 AI leaders, with four layers of industry self-regulation: internal audit, external audit, an industry advisory committee and labs validating each other's models.

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

Watch: David Shapiro: Dots, Bots, and Math! Oh my!

Models and learning

AI Engineer workshop: FP8 and vLLM tuning on Qwen3 4B, speculative decoding slowed demo

In a live load test (about 256 input and 128 output tokens per request), a speaker moved a Qwen3 4B model from BF16 to FP8 and single-stream decode rose from 37.5 to 53.4 tokens/s, with peak throughput about 1.4x higher.

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

Watch: AI Engineer: Agents That Own Their Inference — Du'an Lightfoot & Khaja Omer…

FriendliAI-sponsored talk compares GLM 5.2 with Opus 4.8 on cost and speed

A speaker prompted GLM 5.2 and Claude Opus 4.8 once each to build a tower defense game; both produced working games and he said GLM 5.2 cost more than 5.6 times less.

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

Watch: AI Engineer: Same Model, Different Speed: Why Your Inference Provider Matters —…

Hermes Agent: 16GB Mac Mini setup with Gemma 4 12B and new built-in kanban

Bart Slodyczka ran Hermes Agent on a 16GB Mac Mini with Gemma 4 12B (4-bit MLX, 6.77 GB) through LM Studio, and said Hermes needs a context window of at least 64,000 tokens or it errors.

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

Watch: Bart Slodyczka: The Ultimate Beginner’s Guide to Hermes Agent (Step-by-Step)

Workshop tests agent attacks on Kubernetes: guardrails, Kyverno policy and malicious MCP tool

In a hands-on workshop using kagent, presenters said model refusals were inconsistent: data exfiltration worked at least once then stopped, and a fork bomb never ran.

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

Watch: AI Engineer: Build a Platform and Watch It Burn — Michael Forrester, Accenture &…

Weco says its Aiden agent landed 7 Parameter Golf records

Weco said its multi-agent system Aiden, built on Claude and GPT model families, landed seven merged records on OpenAI's Parameter Golf leaderboard, twice the next best human, and that it plateaued from April 15 to 19 until a community tokenizer improvement was fused in within eight hours.

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

Watch: AI Engineer: Parameter Golf with AutoResearch — Vayum Arora, Zhengyao Jiang…

Typedef claims data context layer raised data-agent task success by almost 35 points

Typedef co-founder Yoni Michael said a deterministic graph of grain, join safety, metric meaning and dependencies raised data-engineering agent task success by almost 35 points in the company's own runs.

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

Watch: AI Engineer: The Data Context Layer for Agents — Yoni Michael & Brandon Callender…

Browserbase presents AutoBrowse doer-teacher agents and browse.sh skill library

Browserbase described AutoBrowse, in which a doer agent drives the browser and a teacher agent distills a reusable skill.

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

Watch: AI Engineer: Hill-Climbing Skills: Improve Agents Without Changing the Model —…

NVIDIA presents Data Designer synthetic data tooling and says new Nemotron Ultra was released

An NVIDIA speaker said the open-source NeMo Data Designer has produced 6.5 trillion tokens of training data, that search-trajectory SFT data moved BrowseComp from 0 to 30 and a text-to-SQL dataset moved BIRD from 26 to 41% for Nemotron Nano, and that NVIDIA just released Nemotron Ultra with two smaller Nemotron models.

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

Watch: AI Engineer: Teaching Agents to Search with NVIDIA Data Designer — Dhruv…

Coverage

Window: videos uploaded in the 24 hours to Oct. 12, 02:14 UTC. Of 410 videos inventoried from monitored channels over the four-day scan, 75 were read for this edition (all 75 from the window; there were no late finds), 283 had been covered in earlier editions, 2 had no transcript available, 11 were live or premiere videos deferred and 50 were under 45 seconds.

Sources reviewed (75 videos)

All 75 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.