Models
Thinking Machines releases Inkling 2 Mini, an open model for laptops
The company says it gets 48.2% on DeepSWE v1.1, a code test, more than all open models with less than 100 billion parameters.
What matters in AI.
SubscribeNews of the day
45 stories filed under this day.
Models
The company says it gets 48.2% on DeepSWE v1.1, a code test, more than all open models with less than 100 billion parameters.
Agents
An agent on Amazon Bedrock AgentCore can send tasks to an SAP Joule agent with the A2A open protocol, in one direction only.
Models
Sol is open to more Copilot users than Astra, which only Frontier users get in Word, Excel and PowerPoint.
Security
Signal stops screenshots by default on iOS and Android, and AI agents in the phone cannot read messages from the screen.
Industry
House committees tell that Chinese open weight models at US companies are a risk that is hard to remove and has a high cost.
Compute
The new limit is 20 times the limit of 2 KB before, and it has no cost.
Deals
No company has confirmed the talks, and the source gives no price or terms.
Compute
The gateway sends each request to the best cluster in each Region, and moves traffic off a cluster that stops.
Security
Amazon rates the vulnerability as a zero-day, one with no patch, with high confidence, as it was in no database.
Industry
Hundreds of Google workers tell leaders to cancel the policy, and leaders did not answer the points of the petition.
Industry
Danaher tells that novices with its AI assistant on a Leica microscope did the task with a maximum of 83% less time in an internal test.
Agents
Google tells that version 4.0 adds an AI agent that finds the cause of a data task error and gives a repair.
Policy
The unit will apply the RAISE Act, which tells makers of large AI models to show safety measures to all persons and report incidents.
Devices
The Dutch Data Protection Authority says the AI glasses that it tested are dangerous for users and for bystanders.
Policy
Bessent says the US policy on H200 chip sales to China is the same, but Nvidia says it stopped them.
Compute
In North America, bookings of circuit boards rose 56.1 percent in the first nine months of 2026, but shipments rose only 13.3 percent.
Models
Claude Fable 5.2 ranks first on three tests, but Artificial Analysis finds the highest cost for each task of all models it priced.
Models
Atlas exports 3D scenes that developers can use in game engines, but most of its test results come from World Labs.
Industry
Shoppers can pay with Klarna installments when an AI agent buys for them, and merchants add it with no new integration.
Security
Zendesk says that personal agents will send about half of all support traffic by 2030.
Agents
Bryna Kra of Northwestern University told that math changed more in two months than in 100 years.
Industry
Cursor made its coding model with a Chinese model. It plans to explore an end to that.
Deals
At Slush in Helsinki, investors said that AI makes software easy to copy. They put money in assets and trust.
Research
Links in the AI summary gave only about 7% of the clicks, and they do not stop the fall.
Industry
George Colville says agent makers must owe users a duty of loyalty, and the right to move to a different agent is not sufficient.
Deals
No company has confirmed the deal, and the price is lower than the earlier value of Inception, reports say.
Industry
Unlimit and Visa will test refunds, chargebacks and safety checks for agent payments, but they gave no results.
Security
Hunt.io tells that a Russian speaking group got bank and government logins from the PCs. No bank or agency was breached.
Security
Signal tells that AI agents in phone and computer systems can read its encrypted chats from the screen.
Research
A check that uses the same data as the writer found only 35% of the added errors.
Research
The correct quantity of agent memory is a property of the model, not of the memory.
Research
But the time from report to repair, and from disclosure to attack, did not change much so far.
Research
When a user sends an image that is not safe with a neutral instruction, Gemini and GPT refuse less than 1 time in 4.
Research
A warning that tells the model about the danger after each tool output cut damage for all 5 models. A note of the same length did not.
Agents
In a test with parallel calls, 21 results went to an incorrect request and 22 were empty.
Research
After 30 rounds of changes, all models had lower results on all benchmarks, by 15 points on average on SkillsBench.
Research
NariQ stays level with the full model at 3.0 and 2.5 bits, where other methods did at most 72% as many long agent tasks.
Security
A cryptographic trace of all steps let an auditor find 98% of attacks and the agent that did each one.
Research
The authors show that no algorithm can know before if an agent will operate a dangerous program, thus no guardrail or certification is complete.
Policy
The White House office that helped write the voluntary AI test plan will get an acting head from the Pentagon.
Security
Enclave 1.0 puts limits on the items that AI coding agents can touch, but its maker tells that it does not make agents fully safe.
Industry
Co-founder Chenyu Zhao tells that a company with an AI cost of $10,000 a month must look at open models.
Security
The package prints a correct report, but on Windows it also starts a program that can get browser passwords and Discord tokens.
Agents
Meta tells that these are isolated incidents, and that Muse uses other apps only with permission that the user gives.
pymnts.comClaimed, not confirmed
Policy
In a joint letter, four UK regulators tell that some developers did not speak to them at the start. They will use their powers.