Daily Briefing
11 November 2026 · 5 stories
🔥 Top story
1Agents prove that Signal's encryption holds
Google DeepMind handed its agents a job that human teams used to spend years on: proving mathematically that Signal's encryption, the app that protects your private messages, really works. The agents wrote more than six million lines of Lean, a language that serves as a calculator for proofs. Experts had to re-read only one percent of the result, the machine checked everything else. The whole job took four weeks. The analogy: it is a full roadworthiness test of a car, bolt by bolt, instead of a glance at the bodywork plan. To make sure the test did not pass by itself, the team put two real bugs that had already been fixed back into a copy of the code, and the system caught both. Source : https://deepmind.google/blog/introducing-alphaproof-code/
2Claude joins your meetings and drafts the plan as you talk
Anthropic adds Claude to video calls, like a visible participant who works while you talk. During the meeting it drafts the plan, maps out a timeline on a whiteboard, answers in the meeting chat and fetches information from the connected work apps. Afterwards the notes, the action items and the documents stay available. Two guardrails are explicit: Claude does not join a call with an outside guest without authorisation, and it leaves as soon as an outside guest joins. The analogy: it is an intern who settles into the meeting, writes on the whiteboard and slips out when a stranger walks in. Anthropic also adds a shared whiteboard and a spreadsheet where Claude writes into the cells, spelling out every change. Source : https://techcrunch.com/2026/11/10/anthropic-launches-claude-meetings-to-turn-workplace-discussions-into-plans-and-tasks/
3One command turns your request into a team of agents
Google opens Gemini 4 Carbon to the public, presented on 28 October but until now reserved for a few cybersecurity specialists. The model is built for long tasks: chaining research, code and verification without giving up at the first obstacle. One command, /swarm, turns a single request into a coordinated team of agents you can follow live. The analogy: it is a site foreman who gets one instruction and sends out crews, with an opposing expert testing every change. The price is 2 dollars per million words read and 10 per million words written, until 30 November; after that date the tariff doubles. Google puts it ahead on several real-work measurements, at a cost per task of less than half that of the best outside model. Source : https://deepmind.google/blog/gemini-4-carbon-is-now-available/
4The box that keeps agents in line becomes a shared standard
NVIDIA hands OpenShell, the software that locks agents inside a constrained space and applies strict rules while they work, to the foundation that hosts Kubernetes. The project becomes an incubation project: it will be governed by the companies that use it in their products, and no longer by a single editor. Two pieces complete the gift: a project to run agents for days on clusters, which keeps their environments between two tasks and frees the machine while they sleep. NVIDIA says hundreds of organisations already use OpenShell, and that the project received tens of thousands of contributions in under a year. The analogy: it is a safety standard moving from an internal prototype to the rule of the sector. Source : https://blogs.nvidia.com/blog/open-source-openshell-infrastructure/
5Google Cloud puts agents in hospitals, without letting them sign
Google Cloud presents Gemini for healthcare, an offer that puts agents to work in hospitals and health insurers. They summarise medical records, prepare appointments, check coverage and protocol compliance, all without ever writing into the patient file: at Epic, Oracle Health and similar providers, reading is the only allowed mode. Launch customers include HCA Healthcare, Northwell Health and Stanford School of Medicine. The figures come from the vendors: nearly 28 hours a week lost to paperwork in a Harris Poll and Google Cloud study, and 41.9 percent of doctors burned out according to the American Medical Association. The analogy: it is a medical secretary who reads everything, checks everything, and signs nothing. Source : https://www.googlecloudpresscorner.com/2026-11-10-Google-Cloud-Launches-Gemini-for-Healthcare
📡 To watch
Norway signs the Paris call on AI safety
Six countries, Norway among them, signed a joint text in Paris on the safety of artificial intelligence. The signatories promise emergency stop mechanisms, testing by independent experts and the sharing of alerts between serious incidents. But the French security organisation, which signed it too, says the text has no teeth: no common risk threshold, no deadline, no power of suspension for the authorities. Google, Meta and China did not sign. Source : https://www.elysee.fr/emmanuel-macron/2026/11/10/9eme-edition-du-forum-de-paris-sur-la-paix
New York opens a registry for the big AI labs
The state of New York opens a registry where large AI labs will have to sign up before 1 January 2027. The law, signed in December 2025, requires publishing safety plans and reporting any serious incident within 72 hours, far faster than the 15 days demanded in California. A new public body, the DIGIT, is created to enforce it, with fines of up to 3 million dollars. The threshold that triggers the obligation: 500 million dollars of annual revenue. Source : https://www.nysenate.gov/legislation/bills/2025/S6953
Anthropic sets its IPO price: 70 to 78 dollars a share
Anthropic has launched the roadshow for its initial public offering and set a range of 70 to 78 dollars a share, read in the document filed with the Securities and Exchange Commission on 10 November. Trading on the Nasdaq, under the symbol ANT, is due to start around 19 November. The banks still wanted to raise the target valuation, an increase that does not appear in the document. Source : https://www.anthropic.com/news/roadshow-launch
OpenAI buys a travel agent and closes the service
OpenAI buys Odessia, a travel agent that prepared flights and hotels, and hires its founder, who already built a large home-rental platform in the same spirit as Airbnb. The agent was built by directing teams of agents and is said to have prepared travel worth several million dollars, according to its founder, a figure nobody has verified. The service was already closed when the deal was announced, and no travel product is planned for now. Source : https://techcrunch.com/2026/11/10/openai-acquires-odessia-travel-startup-created-by-sonder-founder/
📊 Trend
A day on which we saw one agent prove that the encryption of your messages holds, another settle into your meeting to write the plan, and a team form from a single command. Agents are moving from promise to proof. But the rules meant to frame them stay vague: the text signed in Paris sets no threshold, New York demands incident reports within 72 hours without independent checks. Between the demonstration that convinces and the rule that binds, the gap closes slowly. The next step is decided there: agents whose work you can verify, not just believe.