United Airlines has banned outside AI agents from booking its flights, and its CEO says he doesn't want "another middleman." Google released Gemini 4 Hydrogen, a cheaper everyday model that edges past its own Argon on a major ranking. AWS shipped a free tool where no human wrote or reviewed a line of code. Further down, J.Crew gets a Claude-powered stylist, and Notion confirms some AI chat transcripts were taken in a supplier's breach.

IN THIS ISSUE · 5 SECTIONS

 1  Major AI News: United's ban on AI booking agents, Google's Gemini 4 Hydrogen, AWS's agent-built emulator.

 2  Fun AI Topics: J.Crew's Claude stylist, Gemini smart glasses, Gemini on Apple CarPlay.

 3  AI Tools: Gemini Decisions API, Nectar Creator Agent, Fish Agent, Nod, Daybreak Red Team.

 4  Expert Prompt: decide what outside AI agents can and can't do with your customers.

 5  Trending: Notion AI chats taken in a supplier breach, Databricks' always-on Genie, DeepMind's AlphaGenome 2.

LEAD STORY

United Airlines has banned unauthorised third-party AI agents and assistants from its website and app, and from booking United flights through travel agencies and aggregators. The airline gave three reasons. Customers who hand their login to an AI assistant could expose their personal data. An agent could make a purchase, change an account or spend MileagePlus miles nobody meant it to. And automated check-in could break FAA and DOT rules, which need a person to confirm their bags hold no banned items. CEO Scott Kirby told the Wall Street Journal: "I don't want to create another middleman."

Why it matters: United says banks and retailers have taken similar steps, and it's urging other companies to follow. Any business hoping to sell through AI agents now has to plan for big brands that would rather keep the customer to themselves.

ALSO TODAY

Google has released Gemini 4 Hydrogen, which it calls its "daily driver" for agent work, leaving long, deep reasoning to its bigger Carbon and Argon models. It costs $1.50 per million input tokens and $7.50 per million output tokens, at the same rate across its full 1 million-token window. Google says it scores 51 on the Artificial Analysis Intelligence Index at its default setting, one point above Gemini 4 Argon, and finishes tasks 2.5 times faster than Gemini 3.8 Flash. It's rolling out in the Gemini app, Google Search, Gemini in Google Sheets and the Gemini API.

Why it matters: Most everyday AI jobs, like drafting, sorting and filling in spreadsheets, don't need the top model. A faster default that scores close to it makes that work cheaper to run.

ALSO TODAY

AWS has released AWS Emulator, a free tool that runs a stand-in for its cloud on a developer's own computer, so apps can be tested offline without paying for cloud time. AWS says its team hasn't written or reviewed any of the code. Hundreds of AI agents work on it in three teams: one writes tests and checks them against the real AWS, one runs those tests and logs every mismatch, and one fixes the code. It started as a weekend experiment, when an engineer asked an AI coding agent to rebuild one service, Amazon SQS, and got a first version in about 15 minutes.

Why it matters: AWS is handing developers a product built and kept up by agents, with automated security checks where a human reviewer would normally sit. Small teams also get a free way to test against AWS without running up a bill.

J.Crew has launched Style Advisor, an AI stylist on its US website built on Anthropic's Claude. You describe an occasion, like a beach wedding or a first job interview, and get complete outfits back, which you can keep changing or build around one item. It only picks from outfits J.Crew's own stylists put together for catalogue shoots, so every look is one a person already approved. Sonnet 5.5 runs the conversation and the smaller Haiku 5.5 handles quick jobs to keep it fast.

Why it is interesting: Limiting the AI to looks the brand's stylists already made is a simple way to stop it suggesting something off-brand.

Source: The Next Web

Google and Samsung have opened preorders for Gemini smart glasses made with Warby Parker and Gentle Monster. Warby Parker's three styles start at $395, Gentle Monster's six cost $449 to $489, and they ship on 10 December. Say "Hey Gemini" and they can tell you what you're looking at, add an event flyer to your calendar or suggest a recipe from the food in front of you, along with directions, live translation, music and calls. They weigh about 41 grams and last up to nine hours on a charge.

Why it is interesting: The privacy rules are built into the hardware. The camera stops working if you cover its recording light, and other people's faces are blurred before Gemini sees the picture.

Source: The Next Web

Gemini now has its own icon on Apple CarPlay, so iPhone users can talk to it hands-free from the car screen without opening Google Maps first. Google pitches it for preparing for meetings on the commute, going through your email, hearing your Daily Brief and finding places nearby. Chats sync with the Gemini app on your phone, so you can pick one up later.

Why it is interesting: The car is one of the few places people already talk to their phone without looking at it, which makes it a natural home for a voice assistant.

Source: Engadget

[DEVELOPERS] Gemini Decisions API - Answers up to 1,000 yes/no, multiple-choice or scoring questions about one piece of text, image, audio, video or PDF in a single call, with a confidence score on each answer, from $0.10 per million input tokens. blog.google (launch announcement)

[MARKETING] Nectar Creator Agent - Searches more than 300 million Instagram and TikTok creator profiles, sends outreach in your brand's voice and keeps track of gifting, contracts and campaign results. nectarsocial.com (company site)

[VOICE] Fish Agent - Builds an AI phone agent with no code: describe it, give it a number and publish. 3.5 cents a minute at the launch price until 31 December, with $5 of free credit. fish.audio (product page)

[AUTOMATION] Nod - n8n's free, open-source personal agent that works in your apps and websites from a private space on your own Mac or PC, using the AI model you choose. Research preview. blog.n8n.io (launch post)

[SECURITY] Daybreak Red Team - OpenAI's early-access tool where a team of AI agents tests your web app for weaknesses, checks bugs that have been reported to you and tests the fixes. theregister.com (launch coverage)

Context: United's ban shows that any business selling online now has to decide what outside AI agents may do for its customers. This prompt turns that into a one-page policy you can put to work.

Act as the operations lead for [company], which sells [product or service] through [website/app]. Our customers are starting to use AI assistants to [browse / buy / manage their accounts]. Draft a one-page policy for outside AI agents. List every action a customer can take with us, from browsing to paying to changing account details, and mark each one Allow, Allow with customer confirmation, or Block. Give one reason for each choice, tied to a specific risk such as an accidental purchase, a shared password or a legal requirement in [country/industry]. Do not recommend blocking everything: if an action is low risk and helps sales, say so. If you're not sure whether a rule applies to our industry, list it as a question for our lawyer instead of guessing. Finish with a three-line note we could post for customers explaining the policy.

Example use case: A small online wine shop runs it and decides agents can browse and fill a basket, but a person has to confirm their age and the payment at checkout.

Notion has confirmed that hackers took chat transcripts from some Notion AI customers after a breach at Fireworks AI, a company that runs open AI models for other businesses. The transcripts came from Notion's Insights feature, and Notion says about 1% of its customers were affected, with much of the data impossible to tie to a specific customer. Notion cut its connection to Fireworks on 3 November and deleted the data stored there. Fireworks says the attackers seem to want to sell stolen access to AI models, and this is its second breach in two months.

Why it is important: Text typed into an AI feature can pass through suppliers you've never heard of, and a breach at any one of them can expose it.

Source: TechCrunch

Databricks has unveiled a new version of its Genie assistant that runs in the background, follows an office worker's day, makes suggestions and pulls up information before anyone asks. A new Ontology feature maps how a company's data fits together, such as who reports to whom and which team owns which project, so staff can find answers or run an analysis quickly. CEO Ali Ghodsi says he uses it every day. Databricks hasn't said when it will be available or what it will cost.

Why it is important: It's an early look at work assistants that act on their own during the day, instead of waiting to be asked a question.

Source: Fortune

Google DeepMind has released AlphaGenome 2, a model that predicts how changes in DNA affect specific types of human cells. DeepMind says it finds nearly twice as many gene-changing variants as the previous best model, after training on more than 100 million single-cell profiles from 68 tissues. It's available free for non-commercial research, and it isn't approved for clinical or diagnostic use.

Why it is important: Knowing which cells a genetic change affects tells researchers where to start looking for a disease's cause, one of the slowest early steps in drug research.

Source: Google DeepMind

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas