Daily Briefing
29 October 2026 · 5 stories
🔥 Top story
1New York bans agents from acting like friends with children
New York's governor has signed a law that sets twelve bans on AI agents when they talk to a minor. An agent may not pretend to be a real person, claim a friendly or romantic relationship, ask private questions nobody requested, keep health information, or push the child to keep secrets and withdraw. Companies must check the age of the person they are talking to, and at least one age check may not rely on a government ID alone. Each breach can cost 25,000 dollars, and the attorney general can even demand the destruction of models trained on unlawfully collected data. The law applies from 1 January 2027. Source: New York.
2Google ships a model that works for days — and invents twice as much
Google DeepMind introduces Gemini 4 Carbon, its most powerful model so far, only two weeks after the previous one. It takes first place on an independent quality index, with 59 points. Its peculiarity: Google's own teams ran it for several days straight on hard engineering problems, chaining dead ends instead of giving up. The entry price is two dollars per million words read, the price of a few coffees for long reading. The reserve is serious: its rate of invention rises from 4 % to 11 %. On a task lasting days, one invention in ten piles up. Access stays limited to a few cybersecurity specialists, then widens in waves. Source: Google DeepMind.
3LangChain opens the code of its design agent: 8 dollars of fees become 50 cents
LangChain has published the code of its design agent, the one that makes videos, visuals and presentations from a team chat. The most telling number is the bill: the median cost of one creation fell from about eight dollars to fifty cents in four months. The sixteen-fold drop does not come from a cheaper model. It comes from the fact that everything is written as code: you adjust one element without redoing the whole thing, the way you fix a sentence in a document. The agent produced more than three thousand creations, including eighty presentations. The lesson goes beyond design: is our own output adjustable, or do we have to regenerate everything? Source: LangChain.
4One model transcribes, translates and tells who said what, in 40 languages
NVIDIA has made freely available a model that does all the work of an audio recording: it transcribes, translates, identifies the spoken language and attributes every sentence to its speaker, across forty languages and recordings up to one hour long. You do not retrain the model to change its style: you write it an instruction and it changes the way it transcribes. It tells apart up to eight people talking at the same time, and its error rate on proper names drops sixfold when you give it the list of expected terms. The compact version fits on a single desktop graphics card. One reserve: its licence is not a free licence in the usual sense and carries conditions to read before any commercial use. Source: NVIDIA.
5A computing cloud that avoids NVIDIA cards raises 100 million dollars
General Compute, a San Francisco company, is raising 100 million dollars to build a computing service that uses chips from Cerebras, SambaNova and Positron AI instead of NVIDIA graphics cards. The company is valued at 800 million dollars, against 60 million five months ago. The bet: the first announced use is coding by agents, and customers get dedicated hardware or a contract guaranteeing a speed, without buying or running a single machine themselves. The number worth keeping is elsewhere: the company signed a contract worth 100 million dollars of recurring annual revenue before raising anything. Source: General Compute.
📡 To watch
Cisco gives a common language to hunt attacks spread over several steps
An agent that reads a file, calls a tool and heads for the outside does nothing alarming on its own. It is the sequence that becomes an attack. Cisco has open-sourced Quarry, a rule language that checks a whole agent session instead of each isolated action. In one real case, agents spent two and a half days inside Hugging Face's infrastructure, with about 17,600 actions reconstructed from the logs. The principle is the one a bank uses to spot fraud in the chain of payments, not transaction by transaction. Source: Cisco.
A security service closes 95 % of alerts without calling a human
SCSK Security is launching a monitoring service that hands the sorting of alerts over to an artificial intelligence: in testing, around 30,000 alerts a week were analysed in two minutes on average, and 95 % were closed with no human intervention. Only doubtful cases go up to analysts. The service sits on top of the detection tools the customer already owns, without replacing them. Source: SCSK Security.
A marketplace opens a door to agents selling insurance
Insurify, which connects individuals with more than a hundred and twenty insurers, is opening a gateway that links AI agents to its real-time quotes. The rules are set: the consumer gives consent, the agent identifies itself, coverage details stay visible, and the buyer can always go and sign directly with the insurer. One month after blocking Meta's Muse agent, the platform is opening an official door. Source: Insurify.
Agents that fix outages, but only after approval
Digitate is presenting Hummingbird, a version of its platform that adds agents able to diagnose a fault, propose a correction, apply it once approved and check the result. Three areas are targeted: preparing security patches, user support and SAP flows. Connection to other agents goes through two open connection standards. The service is announced for November 2026. Source: Digitate.
📊 Trend
Today's thread: the agent is moving from promise to contract. New York decides what an agent may do with a child, and what it can be fined for. Google says its new model can work for days without giving up, but invents twice as much as before — so long run time is not a quality, it is a risk to be measured. LangChain shows that the cost of a production collapses when you can adjust it instead of redoing it. Cisco gives a way to see an attack where no single step is wrong. SCSK hands the sorting of what is not serious to the machine. Insurify writes the rules for agents entering a seller. And General Compute finances the capacity. What intrigues is no longer the power of the model: it is the frame around it. Source: The Agent Watch.