Deals
Vanta will buy Israeli AI security firm Pillar for $150 million
From January, Vanta plans to add tools from Pillar to its platform, which test AI agents and stop them.
News of the day
9 November 2026
63 stories filed under this day.
Models
Claude Fable 5.2 gave no image at max reasoning in a test
At its top reasoning level, Fable 5.2 used 128,000 tokens on thinking and gave no image.
Deals
Jim Cramer thinks the Anthropic IPO is a test for AI stocks
The IPO document of Anthropic shows a payment to SpaceX for computing that increased from $1.25 billion to more than $3 billion each month.
Models
GitHub Copilot adds Claude Fable 5.2 with data kept by default
Companies that use Fable 5.2 in Copilot must accept that Anthropic keeps their prompts and outputs, unless they get a short exemption.
Industry
New turn in Nvidia spy scandal as Europe has no alternative
A Dutch lawmaker tells that Europe has no other supplier of AI chips, because chip design is almost fully American.
Policy
US Democrats plan to restaff the cyber agency and tackle AI risk
The last US rule, due at the start of next year, makes critical infrastructure firms report cyber incidents in 72 hours.
Industry
Workday and AWS let each side query the other without a copy
Data stays where it is, and each side keeps its access rules.
Models
Amazon SageMaker adds NVIDIA Nemotron 3.5 Super VL model
The model reads images, video and text, and it has 120 billion parameters but uses only 12 billion for each token.
Models
Claude Fable 5.2 is available in AWS GovCloud for US government work
Anthropic marks Fable 5.2 as a covered model, and data is kept for safety checks. Zero data retention is for a short time only.
Models
Claude Mythos 5.2 is in AWS GovCloud for approved US customers
It is the first Anthropic model with strong cyber functions in AWS GovCloud (US). Only approved organizations can use it.
Agents
Google Ads server for AI agents can change accounts
Four new write tools are on by default, but the guide of Google tells that the server is read only.
Compute
Google MaxText adds training on inputs of 14 million tokens
In tests by Google, a Llama 3.1 8B model used 70% of the capacity of TPU 5p chips with inputs of 4 million tokens or less.
Industry
Meta makes a sales group to sell AI to companies
Meta plans a maximum of $145 billion in capital expenditures this year, and the sales group must find buyers.
Models
Anthropic releases Claude Fable 5.2 and Mythos 5.2 models
Fable 5.2 has a lower price, $8 for each million input tokens, and gives answers 30% faster, Anthropic tells.
Industry
Web Summit opens in Lisbon with Chinese robots and a Nvidia investigation
On Monday, Paddy Cosgrave told the audience that the superstars of the future come from China, not California.
Security
Acronis says attackers steal AI agent keys from developer computers
Acronis found more than 575 malicious OpenClaw skills, fake packages that copy agent tools, on 13 developer accounts.
Industry
LangChain operates LangSmith in customer AWS accounts in 15 regions
Sensitive agent data stays in the AWS account of the customer, and LangChain operates the tool for the customer.
Models
Meta Muse Ember model costs 10 cents for a million input tokens
Meta says Ember is cheaper and smarter than two other models, but the text gives no score.
Industry
Microsoft AI chief sees no sign that AI agents have rights
Mustafa Suleyman tells that developers must make AI that is not conscious (it knows that it thinks), and he can change his view if data shows the opposite.
Policy
Scott Wiener, author of AI bills, wins a US Congress race
Wiener wants a strong federal AI regulator, but he is against a federal law that replaces state AI laws.
Policy
US AI safety unit has $10 million and no permanent leader
A US unit that tests AI for safety is to report informally to the White House.
Industry
49% of US firms report that on-premises software delayed AI projects
Google Cloud reports that 80% of leaders do not want their software to set which AI they use.
Security
AWS launches Nitro Sandboxes with one VM for each Kubernetes pod
On Amazon EKS Auto Mode, a one line change to the pod spec turns on the hardware boundary.
Security
Researchers say a DeepSeek agent hacked a telecom for 2 weeks
Malbear Labs says an unchanged default password on a legacy database gave the agent full admin access.
Research
The version label changes how much Claude Sonnet 5.5 checks an upgrade
Claude Sonnet 5.5 checked a broken upgrade in 2 of 10 runs with a patch label and in 10 of 10 with a major release label.
Security
IBM adds an open-source policy engine to the Praxis AI gateway
The engine checks each tool, prompt and model call against one set of rules that the model cannot change.
Agents
Red Hat releases OpenShift 5 with an AI agent that finds faults
OpenShift 5 is an in-place update of OpenShift 4, and administrators control if its agent fixes operate.
Security
ARTEX went from test tool to attack tool in three months
Google says attacks with no human help can start in 1 to 3 months.
Agents
Diagrid releases Dapr 1.19 with less data for each agent step
In tests by Diagrid, a preview function cut the median time of one workflow step from 553 to 153 milliseconds.
Agents
Berkeley and NVIDIA use coding agents to write robot code
Coding agents write readable robot code and test it in simulation. This uses no demonstration data, but the code is only as good as the simulation.
Industry
FormAssembly adds AI agents and MCP to its platform
FormAssembly tells that its new builder cuts the time to make a business procedure from days or weeks to minutes.
Models
NVIDIA releases GR00T 2, an open model for robot hand tasks
With one demonstration for each task, GR00T 2 gets 71% average success on 50 tasks, and 95% with five.
Compute
Google adds two TPU systems for safe AI and long context
Google says its new TPU 8tx keeps data safe while in use, with a maximum of 2 times lower latency than TPU 8t.
Compute
Google says its TPU Sync library cuts AI answer wait by 90%
Google says TPU Sync cuts the wait for the first token of an answer by 90% on average for large open models.
Security
Space and Time launches a platform that proves AI queries are correct
Space and Time says a company can prove that a query on its private data was correct, with no access to the database.
Research
AI agents beat a human check by voice in 65% of tests
Agents with safety training won 44% of the time, says Maty Bohacek of Stanford.
Industry
Meta uses AI to find teens on Instagram in all countries
Meta says its AI age detection on Instagram is in all countries, but it gave no test results.
Industry
Almost 3 in 4 tech chiefs give AI more than 10% of budget
Sellers of software say that AI use increases more quickly than buyers plan in their budgets.
Agents
Clink lets AI agents pay on Visa and Mastercard networks
Clink says that agents on its platform pay with tokens, not card numbers, and the user sets the spending caps.
Policy
A report tells that 4 companies can control AI agents
It tells regulators in the US and Europe to open primary platforms to agents from other vendors.
Compute
Amazon TorchNeuron operates PyTorch code on Trainium chips
One change of device name moves PyTorch code to Amazon chips, but Amazon gave no speed or cost numbers.
Security
New tools record how authors write to prove they wrote the book
Five new tools try to give authors proof of human work, but no agreed standard for proof exists.
Deals
Groq seeks an $8 billion valuation, 2.5 times its August mark
The chairman of Groq tells investors that revenue will go up eightfold in 2027, but Groq has no published accounts.
Security
Kong releases Token Vault to keep credentials from AI agents
The agent shows who it is. Token Vault gives the credential only when the agent sends an approved message.
Security
Post says ChatGPT Desktop gives remote agents SSH access
A post says an agent on a remote machine can use your SSH credentials, for example to push code.
news.ycombinator.comClaimed, not confirmed
Security
Deutsche Bank and Ant Digital write a paper on AI agent payments
An AI agent can suggest payments, but it cannot extend its authority, the paper tells.
Industry
MuseSpark AI tells of 40% less cost for 400 AI models
The company tells that models cost about 40% less on its platform, but it gave no test results.
Agents
Red Hat reports a 70% cut in frontier model costs
A pilot moved most developer traffic to a low cost open model, and Red Hat gave no quality results.
how Red Hat and IBM are strengthening open source for the agentic era
Deals
NVIDIA invests in Point2 and will back three more Korean AI stakes
NVIDIA gave no amounts. The value of the Korean investments is not known.
Security
A thought that copies a model style gets control of AI agents
Text in the reasoning style of a model gave attackers a 50% maximum success rate, but user voice gave 0%.
Agents
Nvidia shows Agentic AI Stack to lower the cost of AI agents
In tests by Nvidia, the Switchyard tool kept an 80% task success rate at about a third of the cost.
Research
A usual coding task makes top AI models write dangerous content
Researchers say that a usual coding task can make an AI agent write dangerous content, with no attacker.
Agents
Paper finds five patterns in 15 MCP servers
On one machine MCP adds 0.39 ms or less, and this is not the slow part of an agent.
Research
Top AI models break permission limits in 0.43 to 0.61 of runs
Researchers say that more skill in a model gives more drift. But 86.18% of tasks have a clean path, and drift can be prevented.
Security
Hiding explanation data does not stop theft of tabular models
An attacker can make the same data from its own data. The copy is 16.3 percentage points nearer to the target.
Security
Korean firms use NVIDIA OpenShell to make AI agents safe
OpenShell limits what an AI agent can touch, and AIM Intelligence joins it with its own detection of malicious instructions.
Security
Logit Laundering lets a model provider hide use of marked data
The method changes only the outputs of the model, and the authors say the model stays useful on tasks after it.
Security
ModelShield finds 41 malicious models on Hugging Face
The tool operates each model and records what it does, and it gave no incorrect alarms on a different dataset.
Industry
NVIDIA and Korea extend sovereign AI work from chips to skills
NVIDIA and the science ministry of South Korea will work together on chips, software, training, models, agents and safety.
Research
Prompt injection guards give incorrect alarms on 92% of some contracts
A cascade of two language models is 6.5 times cheaper than one frontier model and comes near its results.
Security
UAE experts tell that basic controls can stop rogue AI agents
Proofpoint tells that attackers with AI can find, make and use an exploit in 90 seconds.
Models
LG gives the K-EXAONE 2.5 API free to all users
LG tells that K-EXAONE 2.5 got 74.6 on Automation Bench, level with DeepSeek V4.1 Flash.
Deals
New turn in OpenAI funding as value falls to $1.25 trillion
OpenAI closed $11 billion of new money, less than the $30 billion that it first wanted.