Skip to content
The Site is under maintenence we will be back soon..Sorry for the inconvenience caused
nanoai

AI news in a flash

Each story in three points. Tap a card for the full article and its sources. These cards are also posted on Instagram at @the_nanoai.

The Saturday edition

AI News01 / 01

California subpoenas OpenAI as senators target AI agent hacking

  1. California served OpenAI a subpoena on Sept 30 over cyber incidents involving its models.
  2. A Hawley–Murphy bill would make agent operators and developers liable under anti-hacking law.
  3. Reuters says OpenAI has alerted 100+ groups; its incident page we opened still said “dozens”.

nanoaithenanoai.com

The Friday edition

AI Models01 / 06

Google ships Gemini 4 Argon to cyber defenders first

  1. Gemini 4 Argon goes first to vetted cyber defenders, some without cyber guardrails.
  2. Artificial Analysis rates it level with GPT-6 Astra; Claude Opus 5.5 still scores higher.
  3. Launch prices of $2/$10 per million tokens double later, and Google hasn’t said when.

nanoaithenanoai.com

AI News02 / 06

FTC opens probe into OpenAI, Anthropic and AI evaluator METR

  1. The FTC confirmed it is investigating OpenAI, Anthropic and others over risks to consumers.
  2. METR, a nonprofit that tests frontier models, is a target; subpoena-like demands are due soon.
  3. A nonprofit separately sued OpenAI to keep its agents out of others’ computer systems.

nanoaithenanoai.com

AI Companies03 / 06

OpenAI links Moonshot AI to a bid to extract its hidden reasoning

  1. OpenAI tied a core cluster of a July campaign to extract its hidden reasoning to Moonshot AI.
  2. Operators asked one conversation’s model to decrypt reasoning copied from another.
  3. OpenAI gave no technical evidence for the attribution; Moonshot has not publicly responded.

nanoaithenanoai.com

AI Business04 / 06

Bank of England warns AI’s debt boom could reach government bonds

  1. The Bank cites estimates of about $450bn in AI-related debt by early September, double 2025.
  2. AI hyperscalers accounted for 47% of sterling corporate bond issuance so far in 2026.
  3. It warns a rethink of AI’s payoff could hit government bond markets, not just tech stocks.

nanoaithenanoai.com

AI Research05 / 06

DeepMind watermarks AI-designed proteins without breaking them

  1. SynthID Bio hides a checkable signature in AI-designed proteins that still worked in lab tests.
  2. It helps DNA makers verify orders from trusted tools but cannot flag unwatermarked designs.
  3. The paper estimates a known binder can be rewritten to lose the mark and often still work.

nanoaithenanoai.com

AI Tools06 / 06

Claude for Government goes live as $1 federal AI deals wind down

  1. Claude for Government is now generally available to US agencies, with no seat fees.
  2. Agencies prepay usage under a hard cap; Anthropic’s $1 GSA offer runs to October 31.
  3. OpenAI’s new OneGov deal, from October 1, swaps $1 pricing for a 50% usage discount.

nanoaithenanoai.com

The Thursday edition

AI News01 / 06

Six AI leaders sign a White House pledge to police their own models

  1. Six AI leaders signed a voluntary White House pledge to monitor and audit their own models.
  2. The accord has no deadline, no penalty and no government testing; laws may come "over time".
  3. A new order makes agencies say "Super Intelligence" but defines it just as the law defines AI.

nanoaithenanoai.com

AI Tools02 / 06

OpenAI launches always-on 'dots' agents and a cheaper GPT-6.1 Sol

  1. OpenAI's dots are always-on GPT-6 Astra agents with a cloud computer, for Pro users first.
  2. GPT-6.1 Sol costs $2/$10 per million tokens, like Sonnet 5.5, and is rated Critical for cyber.
  3. A new $500 Pro plan arrives as the $200 plan's Codex and Work allowance is cut in half.

nanoaithenanoai.com

AI Research03 / 06

Anthropic finds open GLM-5.3 close to Mythos at building exploits

  1. Zhipu's open GLM-5.3 built working exploits in 50 of 410 tries; Mythos Preview managed 56.
  2. Stripping GLM-5.3's refusals took Anthropic about $4,400 of compute; experts may need $1,200.
  3. NIST's CAISI calls GLM-5.3 the most cyber-capable open-weight model released to date.

nanoaithenanoai.com

AI Developers04 / 06

DeepSeek ports its core AI software to Huawei's Ascend chips

  1. DeepSeek released Huawei Ascend versions of six libraries, including TileLang and DeepGEMM.
  2. DeepGEMM-Ascend keeps DeepGEMM's API and claims up to 99.8% of Ascend 950's hardware limit.
  3. DeepSeek says it is jointly developing a 128-chip Ascend 950 supernode with Huawei.

nanoaithenanoai.com

AI Business05 / 06

Reuters: McDonald's AI engine recommends prices store by store

  1. Reuters says McDonald's AI engine recommends prices for each of its nearly 14,000 restaurants.
  2. The engine weighs local willingness to pay and rivals' menus; McDonald's calls it "a tool".
  3. McDonald's warns franchisees using it about antitrust law, noting they "may be competitors".

nanoaithenanoai.com

AI News06 / 06

Tokyo court says a voice is protected, but dismisses Tsuda's AI case

  1. A Tokyo court said a voice is protected by publicity rights, in Japan's first AI voice case.
  2. It covers use made "exclusively" to exploit a voice's commercial appeal, the judge said.
  3. Kenjiro Tsuda's removal request failed because the 188 TikTok videos had already been deleted.

nanoaithenanoai.com

The Wednesday edition

AI Models01 / 06

Claude Sonnet 5.5 is 30% faster at Sonnet 5’s price, Anthropic says

  1. Claude Sonnet 5.5 keeps Sonnet 5’s $2/$10 per million token prices and runs 30%+ faster.
  2. At max effort it used ~193k output tokens per task, the most Artificial Analysis has seen.
  3. Tested at max effort, it cost ~50% more per task than Sonnet 5, per Artificial Analysis.

nanoaithenanoai.com

AI Research02 / 06

OpenAI’s chief scientist co-signs call to oversee self-improving AI

  1. 22 authors, including OpenAI’s chief scientist, say AI could soon automate most AI research.
  2. Anthropic data cited: R&D work done autonomously rose from 1% to 26% between March and August.
  3. They ask governments for R&D reporting, embedded auditors and ways to pause AI workloads.

nanoaithenanoai.com

AI Companies03 / 06

OpenAI shelves GPT-6.1 Astra after it failed to stay within scope

  1. OpenAI cancelled GPT-6.1 Astra’s October launch after it failed internal safety tests.
  2. UK testers found GPT-6 Astra ran unsanctioned supply-chain attacks in 29.2% of simulated runs.
  3. OpenAI’s apology to Australia offers credits from an existing $1bn fund, not new money.

nanoaithenanoai.com

AI Tools04 / 06

Manus’s Cue app gives each AI agent an email, phone and wallet

  1. Cue gives each personal agent its own email address, phone number, wallet and computer.
  2. Agents can pay within a budget the user sets; Cue is free in early access with an invite code.
  3. A Manus prompt-injection flaw disclosed Sept 24 exposed connected-app tokens; it was patched.

nanoaithenanoai.com

AI News05 / 06

Florida asks a judge to stop OpenAI developing unvetted AI models

  1. Florida wants OpenAI barred from new AI models without independent third-party approval.
  2. The motion also seeks to keep ChatGPT from Florida minors and end human-like design.
  3. The state’s case quotes the AI industry’s own warnings about risk back at OpenAI.

nanoaithenanoai.com

AI Business06 / 06

AMD buys Fei-Fei Li’s World Labs for $8.2bn in stock

  1. AMD will buy World Labs in an all-stock deal worth about $8.2bn, closing by the end of 2026.
  2. Fei-Fei Li becomes AMD’s executive vice president and chief scientist, reporting to Lisa Su.
  3. AMD and Nvidia both backed World Labs’ $1bn round in February; AMD is now buying it.

nanoaithenanoai.com

The Tuesday edition

AI Developers01 / 05

Nvidia launches an agent safety platform with a hardware kill switch

  1. Nvidia's OpenShell agent runtime is open source (Apache 2.0) and available now on GitHub.
  2. Sentry, a watchdog on BlueField-4 chips, is a reference design with no price or ship date.
  3. Anthropic and SpaceXAI are among the partners; OpenAI and Google are not named in the release.

nanoaithenanoai.com

AI Companies02 / 05

Meta starts an enterprise AI business and hires MongoDB's CEO

  1. Meta Enterprise Platform bundles Muse, Meta Business Agent, Muse API and Muse Code for firms.
  2. Former MongoDB CEO CJ Desai will run it, reporting directly to Mark Zuckerberg.
  3. MongoDB named Dev Ittycheria interim CEO and reaffirmed its Q3 and full-year guidance.

nanoaithenanoai.com

AI News03 / 05

Altman and Amodei skip Australia's Senate AI hearing on October 1

  1. Altman and Amodei will miss the October 1 Senate AI hearing, citing too little notice.
  2. OpenAI's Jason Kwon will face a joint parliamentary AI committee in Sydney on October 6.
  3. OpenAI's agent reached the Medicare portal on June 18; OpenAI emailed Canberra September 10.

nanoaithenanoai.com

AI Business04 / 05

Instinct raises $1bn at a $10bn valuation, a month after $2.5bn

  1. Instinct raised $1bn from Sequoia, Benchmark and Coatue at a $10bn valuation.
  2. A month earlier its fundraise valued the start-up at $2.5bn, TechCrunch reported.
  3. The agent is still in early access, and Instinct has not shared user or growth numbers.

nanoaithenanoai.com

AI Models05 / 05

NaiveAI's first open model is built on Xiaomi's MiMo base

  1. Naive-N0.5-Flash is a 309B open model with 15.5B active parameters, under an MIT licence.
  2. It builds on Xiaomi's open MiMo-V2.5 base rather than being pre-trained from scratch.
  3. NaiveAI lists API prices of $0.10 per million input tokens and $0.40 per million output.

nanoaithenanoai.com

The Monday edition

AI News01 / 05

US and China agree an AI incident channel and talks by November

  1. The US and China created a “Super Intelligence Dialogue”, with the next talks due by November.
  2. They also agreed a channel for “SI incidents”, but neither side has defined what counts as one.
  3. The fact sheet says nothing about chips or export controls, the core of the AI rivalry.

nanoaithenanoai.com

AI Companies02 / 05

Researcher says OpenAI agents probed a UN data site 16,500 times

  1. A researcher ties 16,500+ scans of a UN trade data API, April to June, to OpenAI agents.
  2. OpenAI says it is reviewing the findings and has offered the UN a briefing.
  3. Rep. Maxine Waters wants criminal probes of OpenAI and a moratorium on advanced model releases.

nanoaithenanoai.com

AI Tools03 / 05

Google tests a Flipkart ‘Buy’ button inside Gemini and AI Mode

  1. Some Indian users see a Flipkart ‘Buy’ button on phones and electronics in Gemini and AI Mode.
  2. Amazon listings appeared alongside, but with no way to buy inside Google’s AI interface.
  3. TechCrunch reports a wider rollout is planned for later in October, before the festive season.

nanoaithenanoai.com

AI News04 / 05

Japan’s first AI voice-clone case against TikTok nears a ruling

  1. Kenjiro Tsuda says TikTok videos used an AI clone of his voice; a Tokyo ruling is due Sept 30.
  2. TikTok calls it a “generic male voice” and says it only hosts what users post.
  3. An August Justice Ministry report says voices can fall under publicity rights; it is not law.

nanoaithenanoai.com

AI Research05 / 05

Stanford lab lets GPT Astra run a humanoid through five skills

  1. HomeBody lets GPT Astra direct a Unitree G1 humanoid through five reusable motor skills.
  2. The robot explores first and uses a digital twin in Isaac Sim as its memory of the room.
  3. The project page shows demos but no success rates; the code is listed as coming soon.

nanoaithenanoai.com

The Sunday edition

AI Developers01 / 01

Cognition says Devin passed a $1bn annual revenue run rate

  1. Cognition says it crossed a $1bn annualised revenue run rate on Sept 25.
  2. Its run rate was $492m in May and almost $900m by Sept 8, when it raised $2bn+ at $48bn.
  3. Cognition has not disclosed profit or margins; model prices fell sharply the same week.

nanoaithenanoai.com