The AI Times
EN
Volume I · Daily editionDaily roundup of AI newsMCPMD
EN
THE AI TIMES
Editorial

AI is getting the keys. Who stays in control?

AI assistants are moving from answering questions to accessing accounts and work history. Today’s selection shows different sides of the same shift: Google is adding conversation exports for administrators, 1Password wants to protect passwords, and banks are restricting some assistants. The promise of convenience therefore comes with the question of who can access our data and what they can do with it.

Gemini · Privacy22 October

Work conversations with Gemini will support targeted exports

Google is expanding its corporate archive Vault to include conversations with Gemini in Workspace apps. Administrators can select specific accounts and periods and retrieve conversation content.

Gemini conversation exports · scope and timelineGoogle · official announcement
Illustration based on the facts in the report

The export includes prompts, AI responses, generated attachments and links to work files used. It covers conversations created with history enabled; users have no separate setting for this feature. Retention and legal hold rules for Gemini in Workspace are due in later releases. Administrators can already control history and deletion at the organizational unit level, including preventing users from deleting conversations manually.

Google Workspace Updates

Muse · Privacy22 October

Muse restores sharing after sensitive conversations were exposed

Meta has restored conversation sharing in Muse, according to The Verge. Researchers had previously warned that public conversation links were easy to find.

Researchers at Claroc reported that shared conversation addresses were derived from their titles and recorded in public DNS, the internet’s address system. The pages found included an internal revenue overview from a marketing company. The researchers also found a small business owner’s loan application.

The Verge AI

Canva · Images22 October

Canva Craft plans a design before drawing it

Canva is introducing its own image generation model, Canva Craft. According to The Verge, it first plans the design and then creates the image, which is intended to produce more accurate and controllable results.

This is a two-stage image generation process; greater accuracy and consistency remain promises from the vendor as reported by the outlet. TechCrunch writes that Canva is making the new image model available to all users.

The Verge AI · TechCrunch AI

AI assistants · Accounts22 October

1Password aims to log AI assistants in without revealing passwords

According to 9to5Google, 1Password offers AI assistants access to accounts without directly seeing the user’s password. Users are meant to retain control over access.

The feature targets assistants that perform web tasks on a person’s behalf, such as making bookings or purchases, as Engadget describes. Engadget names Muse, Dots and Grok Bot as examples of these assistants.

9to5Google · Engadget · TechCrunch AI

AI assistants · Finance22 October

US banks restrict AI assistants’ access to accounts

According to The Decoder, customers are letting AI assistants use their login credentials for major US banks. Banks are responding by blocking or restricting access.

The report mentions Morgan Stanley and JPMorgan; this involves access to real customer accounts, rather than simply talking to a banking chatbot.

The Decoder

AI · Learning22 October

Czech AI tutor Artie secures funding for further development

The Artie app listens to users play and continuously adapts their music lessons accordingly. According to Lupa, its creator ArtMaster has raised another USD 2.4 million, over CZK 50 million, for development.

For people interested in playing, feedback based on their own performance is what matters; the new development is funding for further work on the app. Earlier investors include Robert Kyncl, CEO of Warner Music Group.

Lupa.cz

SOURCElupa.cz
AI · Health22 October

Microsoft reports more accurate diagnoses in a test on selected cases

According to The Verge, Microsoft claims its diagnostic AI achieved up to four times the accuracy of doctors in a test. The report describes a comparison using selected cases.

For ordinary users, it is important to distinguish a test result from use in their own treatment; the comparison described does not itself establish that such a service is available. According to The Verge, the results were published in Science on Thursday. The scenarios were based on medical case reports and emergency department data.

The Verge AI

What experts say

Ethan Mollick

professor at Wharton, author of One Useful Thing

AI handles more work; people still judge the result

Ethan Mollick describes the shift from chatting with AI to assigning work to autonomous assistants. He had the article’s graphics prepared in Claude Projects using Opus 5.5 and chose from the proposed options. However, he warns that this way of working costs us some control over how AI reached its result.

oneusefulthing.org
What experts say

Simon Willison

developer, creator of LLM and Datasette

An AI-created game also needs to be played and refined

Simon Willison describes creating Battle Drones largely with Claude Code. After 12 major revisions, he has a game for mobile and desktop; he barely reviewed the code but played it repeatedly. The example shows the difference between creating a working game and refining an experience that players enjoy.

simonwillison.net
AI · Voice22 October

Amazon Polly adds more natural voices without raising prices

Amazon has upgraded Polly, its text-to-speech service. It promises listeners more natural delivery and more accurate reading of numbers, currencies and addresses; these are the vendor’s claims.

Polly generative voices · current offeringAmazon · vendor data
Illustration based on the facts in the report

The generative part of the service now offers 57 voices in 11 languages and 25 local variants. The update is generally available and, according to Amazon, does not increase prices. It adds 14 voices in 9 local variants. Hindi and Finnish are generally available in this part of Polly for the first time.

AWS ML Blog

Meta · Data22 October

Meta’s terms update fails to ease concerns over using websites for AI

Meta updated its commercial terms and removed wording suggesting unrestricted use of news websites to train AI. According to The Verge, legal experts still see room for content collection.

This also affects people and companies publishing their own work online: according to the lawyer quoted, changing the wording alone does not resolve the scope of possible collection.

The Verge AI

Replit · Apps22 October

Replit gives apps an AI assistant for fixes in production

Replit has launched Monitoring Agent, an AI assistant for running deployed apps. It aims to spare creators of smaller services the work of finding the causes of problems reported by users.

According to Replit, the assistant investigates alerts, identifies the cause, prepares and verifies a fix, and deploys it to production. It then explains what happened and what it changed. It builds on the earlier App Monitoring, which tracked app health and sent alerts. Replit is presenting it primarily to internal app creators and entrepreneurs.

Replit blog

Claude · AI assistants22 October

Claude Duet works even when the user is away

According to The Decoder, Anthropic has introduced Claude Duet, an AI assistant that handles tasks independently even when the user is offline.

According to the outlet, the app requests approval before sensitive actions such as payments. TechCrunch reports that access is through a waitlist for Pro and Max plan users. According to The Decoder, the app is built on Claude Opus 5.5. It connects to services such as Gmail, Slack and Stripe through more than 6,000 connectors. According to The Decoder, Anthropic is responding to similar products from xAI, Meta and OpenAI.

The Decoder · TechCrunch AI · 9to5Google

VerifiedMyth of the day

“Images shared on X show real rituals on Epstein’s island.”
False

Lead Stories checked four images from a video: SynthID detected a watermark for content created or edited with Google AI in all of them. The post linked them to October 2019, although Stephen Hawking, who appears in them, had died in March 2018. The outlet found no matching records in the searchable database of Epstein files.

Verified by: leadstories.com

Repo of the dayGitHub Trending

cloudflare/security-⁠audit-⁠skill

security-⁠audit

An extension for AI coding assistants organizes security reviews of source code. It suits developers who want to verify findings independently and separate confirmed issues from unverified suspicions.

Total stars
39 565
Forks
2 465
Language
JavaScript
What it does
  • Maps the project’s architecture, inputs and trust boundaries.
  • Has a fresh verifier challenge every suspected finding.
  • Creates structured records and reports from verified results.
Get started
npx skills add https://github.com/cloudflare/security-audit-skill \ --skill security-audit

Repository on GitHub →

From ScoutWorth testing

Claude Code

Claude Mods changes when and how AI asks for attention

The author describes how he uses Claude Mods to customize Claude Code’s behavior and the way AI addresses him while he works.

  1. A mod is an extension made up of TypeScript or JavaScript functions that run inside Claude Code. It changes both the behavior and appearance of the command-line and desktop environments.
  2. Extensions respond to prompts, tool use or interface rendering. They can observe an event, change how it is displayed or take over its handling.
  3. The author calls this layer an interface for human attention: it decides when AI interrupts a person, what it shows them and which decision it needs from them.
  4. Claude can help create extensions thanks to a built-in skill, a set of instructions for this work. This lets people work with AI to modify the very way they collaborate.

latent.space · Dan McAteer · Author’s claim

Events in CzechiaNext 14 days · until 5 November

  1. AI models in industry: transformation from different perspectives

    The workshop focuses on deploying and using AI models in industry, combining theoretical insights with a practical demonstration of deployment on supercomputers and addressing the risks of vendor lock-in and other limitations.

    WorkshopOstravaprice on website
  2. Proto Robot #0

    PROTO ROBOT #0 is the first gathering of people working in robotics and physical AI in Prague.

    Česká asociace AI

    Prahaprice on website
  3. AI-Native Company Roundtable

    If you could rebuild your company from scratch today, would you design it the same way?

    DiscussionPrahaprice on website
  4. AI ROI Roundtable

    AI ROI Roundtable will open a practical discussion about the economics of AI transformation.

    Panel discussionPrahaprice on website
  5. Dny AI

    Dny AI is the largest multidisciplinary artificial intelligence festival in Czechia.

    FestivalČeská republika (po celé ČR)
  6. What do machines think with? - Opening of Dny AI in Brno

    The evening opens Dny AI in Brno in cooperation with the Czech Semiconductor Centre.

    MeetupBrnoprice on website
  7. Data Festival VŠE 2026

    Join us for the 9th edition of an all-day event for everyone who loves data, AI and modern technology!

    Česká asociace AI

    Prahaprice on website
  8. AI for Business

    A seminar on assessing your company’s readiness for AI.

    SeminarBrnoprice on website

Model leaderboardIntelligence Index · top 10

Which models are the smartest today

Artificial Analysis combines several independent benchmarks into one number. A higher score means the model handles harder tasks.

  1. 58Claude Opus 5.5
  2. 56Claude Sonnet 5.5
  3. 56Muse Wonder Full
  4. 53Claude Fable 5.1
  5. 53GPT-⁠6 Astra
  6. 53Gemini 4 Argon
  7. 52Muse Wonder
  8. 52GPT-⁠6.1 Sol
  9. 51Claude Opus 5
  10. 50Claude Fable 5

Cost per completed task

  1. Claude Opus 5.55,98 $
  2. Claude Sonnet 5.55,46 $
  3. Claude Fable 5.17,63 $
  4. GPT-⁠6 Astra3,26 $
  5. Gemini 4 Argon1,99 $
  6. Muse Wonder4,52 $
  7. GPT-⁠6.1 Sol0,72 $
  8. Grok 4.73,74 $
  9. MiMo-⁠V2.6-⁠Pro0,13 $
  10. Qwen3.8 Max5,41 $

Best image models

  1. GPT Image 2.5 Sunburst (max)1222
  2. GPT Image 2.5 Flare (max)1207
  3. GPT Image 2 (high)1186
  4. Nano Banana 2.11176
  5. MAI-⁠Image-⁠2.61166
  6. Grok Imagine Image 2.01166
  7. Nano Banana 2 (Gemini 3.1 Flash Image)1140
  8. Muse Image1127
  9. GPT Image 1.5 (high)1123
  10. MAI-⁠Image-⁠2.6-⁠Flash1119

Data: Artificial Analysis · Intelligence Index, higher is better, cost per completed task (Intelligence Index task average) and Text to Image Arena · as of Oct 23, 2026