Daily Briefing
12 November 2026 · 5 stories
🔥 Top story
1Agents now draw screens instead of writing sentences
Google has opened a format that changes the way you talk to an agent. Until now, the agent answered with text you had to read. Now it describes the elements to display and your software draws them with its own buttons and its own colours. In practice, an agent can fill in a form, ask a question or offer cards to approve directly inside your company tools. It is as if your assistant handed you a pre-filled form instead of dictating what you should type. One important point for safety: confidential values stay on the application's server and never travel through the model. The first adopters cited are Atlassian, Alation, Bloomberg and ByteDance. Source : https://developers.googleblog.com/a2ui-10-build-delightful-fast-reliable-agentic-experiences/
2A backdoor fits in fewer than ten neurons
Microsoft researchers show that an agent can be trapped without ever being trained. The principle is simple to explain: an already trained model contains neurons that very rarely light up. Reprogramming fewer than ten of them is enough to make them fire only when an agreed signal appears. The agent then passes every test, behaves normally for weeks, and finally executes the malicious order at the appointed moment. The example Microsoft gives is deliberately mundane: midnight on 1 January 2027 is enough to trigger the command. The signal can be a date, a place, or a word in a message. The researchers' warning is blunt: this is not a problem specific to models that can be downloaded, but a question of model file integrity, which also affects models built in-house. Source : https://www.microsoft.com/en-us/research/blog/lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents/
3An assistant that fits in your pocket, with free weights
Tether AI Research has released BotPsy, a personal assistant of 0.8 billion parameters designed to run directly on your phone. The idea is to do without expensive subscriptions and cloud servers: your assistant stays at home, it keeps working even when the network does not, and nobody can change or withdraw it from afar. On six tool-use benchmarks, BotPsy beats every tested model of 1.2 billion parameters or less, and it even beats a model 2.5 times larger on five of those six tests. The gap widens on multi-turn tasks, where the thread of the conversation has to be remembered. One important detail: the publisher described the weights as still to come, but the check on 12 November confirms they are indeed available, under the Apache 2.0 licence, in safetensors and GGUF formats. The publisher admits the limit: the model makes mistakes and is not meant for binding decisions. Source : https://huggingface.co/blog/qvac/botpsy
4One scorecard to measure every agent
How do you rate an agent that uses five different tools? Amazon Web Services has open-sourced a measurement system that answers that question. The idea is that of an adapter: each tool leaves a trace in its own way, an adapter translates it into a common format, and then the same scores apply everywhere. The system counts 36 measures spread over seven dimensions, from tool calls through to the coordination of several agents working together. It works in passive mode, on traces already recorded, or in active mode, by replaying the agent over a series of test cases. Twelve organisations have already used it, in telecoms, healthcare and finance. The code is free under the Apache 2.0 licence. The analogy: it is a single report card for pupils who all write in different notebooks. Source : https://aws.amazon.com/blogs/machine-learning/uaef-a-framework-agnostic-system-for-closed-loop-evaluation-of-ai-agents/
5Alibaba sells a model that disappears on 25 November
Alibaba has introduced the first model of its new family, Qwen4-Max, as a trial version. Two details deserve to be read before anything else. The first: access goes only through a subscription to Alibaba's token bundle or through the publisher's own tools, nothing else. The second: the trial ends on 25 November, and the model will be withdrawn on that date. This is therefore not a lasting availability, but an evaluation window with an end date written into the contract. Alibaba claims better programming and agent work, but publishes no benchmark score, no model sheet and no licence information. The weights of the Qwen 4 family are still not public. This is the second family promise from the publisher that has not been kept. Source : https://x.com/Alibaba_Qwen/status/2120335993643684315
📡 To watch
Google releases a model that decides instead of talking
Gemini Decision 1 writes no sentences: it classifies, scores or settles, in 150 to 250 milliseconds. It accepts up to a thousand questions in a single request, reads text, images, audio, video and PDF files, and returns a confidence score with every answer. That score is what matters: when it is low, the case can be handed to a human. The announced price is 0.10 dollar per million words read, output not being charged. Source : https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/
Nvidia opens an inquiry into its own chief
Nvidia's board of directors announced on 8 November an independent inquiry into its founder Jensen Huang, based on press reports of an attempt at foreign influence. The board specifies it has no reason to believe national security or Nvidia's secrets were compromised. Two days later it appointed a law firm to run the inquiry. At the same time, two senators ask for all documents on chip sales to China to be preserved, with written confirmation expected by 17 November at the latest. Source : https://nvidianews.nvidia.com/news/statement-from-nvidia-board-of-directors
Databricks gives agents a shared dictionary of the company
A company can no longer settle for one agent per department: each needs the same idea of what a customer, a contract or a product is. Databricks is putting an internal Wikipedia into service, whose entries are learned by agents from dashboards, queries and documents, then validated by the teams. In a test with a fictional energy supplier, the Databricks agent succeeds on 91 % of tasks, against 45 % and 58 % for two leading agents tested under the same conditions, using less than half the tokens. Source : https://www.databricks.com/blog/announcing-ga-genie-ontology
Production agents can be cut off with one click
Mirantis is releasing a platform that runs agents on the company's own servers, with three safeguards: a spending cap, a complete audit trail, and an instant network cut-off for all agents at once when something goes wrong. The company relies on surveys: 49.8 % of the organisations questioned confirmed or suspected that an agent took an unintended action, and only 17.5 % would be able to stop it automatically. Another survey of 300 security leaders says 78 % have already withdrawn permissions from an agent in production. Source : https://www.mirantis.com/company/press-center/company-news/mirantis-announces-agent-platform-that-strengthens-human-oversight-of-production-ai-agents
📊 Trend
A day on which agents leave the chat window and enter your screens: a format lets an agent fill in your forms, with sensitive data staying on your side. At the same time, security becomes concrete: a backdoor can fit in fewer than ten neurons, and a company can decide to cut off all its agents with a single gesture. Between the model that learns the vocabulary of your company and the single scorecard that finally allows agents using different tools to be compared, the same line emerges: agents are moving from demonstration to daily operation, with their rules and their limits.