Policy
New turn in OpenAI copyright lawsuit as judge allows 56 briefs
The judge will give a ruling on fair use of AI training, with no trial.
News of the day
11 November 2026
48 stories filed under this day.
Models
Amazon adds 3 open DeepSeek V4 models to SageMaker JumpStart
The model DeepSeek-V4-Pro has 1.6 trillion parameters, but uses only 49 billion for each token.
Models
Google puts Gemini Live voice chat on Apple CarPlay for all iOS users
Gemini on CarPlay can read your Gmail and Calendar by voice, and the chats sync to your iPhone.
Security
ARTEX attacks can give cloud access in less than a day
The attackers use no software fault: they use public password reset pages and open APIs (connections between software parts) to get cloud access.
Security
AWS patches MCP gateway vulnerability that opens internal services
Versions 1.0.3 to 1.34.0 let a user make the gateway send messages to internal services, and version 1.35.0 repairs this.
Server-side request forgery in the server registration and update API in mcp-gateway-registry
Security
New turn in Cisco BroadWorks attack as CISA sets 24 November deadline
CISA tells US civilian agencies to patch the Cisco BroadWorks vulnerability by 24 November, but Cisco made the patch on 19 August.
Industry
Cloudflare CEO says AI agents will become the primary users of the web
Cloudflare data shows bot traffic was more than human traffic in 2026. The CEO says agents do not click ads.
Policy
Trump accepts AI guardrails but not a slower race with China
Trump named no rule that he will use, and he said China has no comparable limits.
Industry
US firms pay 1.65 percent more for AI in October
The top one percent of firms pay a median $10,200 for each worker each month. A typical firm pays $16.32.
Agents
AWS Lambda MicroVMs can operate x86 software with no change
Teams that use x86 software can put it in a MicroVM as it is, with no change.
Models
Google makes a Decisions API with a result in 150 to 250 ms
The price is $0.10 for 1M input tokens and $0 for output. Google tells that the cost is almost 5 times lower than Gemini 3.5 Flash-Lite.
Security
LLM 0.37.2 repairs two conditions that send an OpenAI key in error
Two conditions could send your OpenAI API key to a different provider if you use LLM with OpenAI-compatible endpoints.
Models
Radical Numerics makes an offline tool for gene threats
The tool gives each gene sequence a threat score from 0 to 1, but the company gave no test results.
Agents
Google releases A2UI 1.0 for AI agents to show screens, not text
An agent can answer with a form or a chart that the app renders, not with long text.
Security
AI assistants change tool servers on their own with admin access
In a lab, 19 makes of one pinned tool server gave 15 different environments in five weeks.
Industry
Amazon SageMaker adds billing by use for live voice AI
Model providers can report the audio that customers use in a live session, and Deepgram is the first partner.
Security
Attackers host AntiFox AI agents on Alibaba Cloud to hit AWS
Arctic Wolf says attackers used AntiFox, an open-source AI attack tool, to get AWS secrets.
Agents
Databricks releases Genie Ontology, a shared layer for agents
Databricks says its agent was correct in 91% of runs, 1.6 times the rate of the stronger rival agent.
Security
Free site gives AI-made exploits for 200 vulnerabilities on patch day
Anyday Security says 72% of the exploits that its AI made had no public version before.
Industry
Helsing shows Clarity, an AI platform that prepares military targets
Clarity prepares targets, and Helsing tells that a person must give approval for each strike.
Agents
LangChain 1.5.0 drops Python 3.10 and adds a model router
Teams on Python 3.10 stay on LangChain 1.4.x and do not get the new release.
Models
Runway releases the Praxis-1 robot model with open weights
Runway tells that Praxis-1 got 46.2% on a robot test, and all developers can use it, also in a company.
Policy
Tallinn summit hears that Europe is slow while AI agents move fast
At the Tallinn summit on 5 November, speakers told that Europe is slow, and they did not agree on how to test AI agents.
Industry
Duffel and Katanox make a partnership to cut hotel booking steps
Bank payments replace virtual credit cards, and a platform can hold funds before the hotel gets the money.
Policy
New turn in Kenya High Court petition against Meta about Ethiopia
On 12 November, the court hears Meta against a petitioner plan to add new proof.
Deals
Plurai gets $13 million to test AI agents in simulations
Plurai tells that small models do the work of large models at lower cost, but it gave no numbers.
Industry
Red Hat releases AI 3.4 for AI agents on hybrid cloud
Red Hat tells that AI agents will cause more use of inference, the operation of models to give answers.
Security
Trust3 AI releases an AI agent kill switch with a trust score
A trust score out of 10 for each agent can start the stop with no reviewer.
Agents
Rain lets AI agents pay from a vault with one-task cards
The agent can spend only the money in the vault, and each card is for one task.
Industry
Dropbox adds an AI agent that plans and runs file tasks
Dropbox says its agent shows a plan first and does the work only after approval.
Industry
ByteDance prepares an AI agent for release in weeks, Bloomberg reports
ByteDance is to release an AI agent of the Muse-type in weeks, but its co-founder tells staff it is far from the top until 2029.
Industry
Phase names the former Cointelegraph chief CEO of its AI agent chain
AI agents pay on it with no human approval, but the funds cannot go out and one funder backs it.
Security
New Citrus ransomware group gets in through contractors and uses LLMs
The group uses stolen contractor logins to get in. Then it uses AI models to write the commands and scripts for the attack.
Models
Fingerprint launches Fraud Score, which caught more fraud in one test
In one test, the tool caught almost 50% more fraud than the model that one customer used before. It did not stop more good users.
Deals
Anthropic files IPO at $70 to $78 as OpenAI loss is $10.9 billion
Seven tech companies agree to buy $26 billion of Anthropic shares at the IPO price.
Research
In 72% of bad AI agent runs, proof of the error was ignored
Researchers say errors start in the first part of a task, and no agent used the proof in 88% of those runs.
Policy
New turn in Australia inquiry on OpenAI agents and medical records
The Australian government will look at a crime for AI companies that release dangerous models. This will not come fast.
Benchmarks
Best AI agent scores 0.337 on a new health evidence test
In a test with no leaked answers, the best of 8 top AI agents got a factual F1 score of only 0.337.
Research
Deleting AI reasoning history shifts task scores by up to 51 points
In a test of 7 open models, the same deletion of reasoning history changed task scores by minus 4.4 to plus 51.3 points.
Policy
New turn in US and China AI dialogue as Bessent names its aims
The US Treasury Secretary said both nations set up a dialogue on "super intelligence". He made no promise to limit AI.
Research
Claude Opus 5.5 leads a new 100 task coding test with 67.3%
Claude Opus 5.5 got 67.3% and GPT-6 Astra got 63.7%, and the difference is not statistically important.
Benchmarks
Deepfake detectors that read a pulse signal learn the video source
Simple color traces got a score of 0.998 on talking face fakes, but the signal is not a pulse.
Benchmarks
Failure labels for AI reasoning change with the sequence of the tests
In all six test sequences, the share of one failure type was between 26.4% and 76.4%.
Research
Test of 17 AI moderators finds no model strong on all five trust axes
In AI content moderators, higher detection accuracy frequently comes with greater demographic disparity.
Models
Tether AI Research presents BotPsy, a 0.8B model for on device agents
Tether gives scores that show BotPsy raises the tool use score BFCL v4 from 23.4 to 44.0 against its initial model.
Deals
Finite State gets $27 million to find flaws in device software
The money will make more AI tools that test the software in complete devices, as new laws put proof on makers.
Security
A report tells that many AI agents work with no name for customers
In one week of October 2026, one in seven of 250,000 agent visits did a task for a customer.
Agents
An MCP server used 42 times the tokens of a CLI on one task
On Claude Sonnet 5.5, one GitHub task cost $127 for 1,000 tasks with MCP and $7.50 with a CLI.