Models
Edge0 releases an open 145B model that operates on a laptop
The Edge0 team reports more than 80 tokens each second on a laptop. Only tests by the team show this.
News of the day
12 November 2026
55 stories filed under this day.
Industry
Microsoft will stop Purview Data Governance support in 2027
Governance moves into Microsoft Fabric without a meter charge, and Purview keeps only data security and compliance work.
Deals
New turn in Nscale share sale as its loss increases to $1.69 billion
Nscale had a loss of $1.69 billion in the third quarter, and it did not tell how much it wants to get.
Security
Nvidia tells that an isolated unit must monitor AI agents
Nvidia tells that its BlueField-4 unit monitors agents from an external position, and agents cannot change it.
Compute
Analysis finds signs of Blackwell chips in 6 Chinese AI models
The analysis says random variation gives about 8 sigma at most, but all its tests show 20 sigma or more.
Deals
IBM buys AuthMind to monitor how AI agents use credentials
IBM plans to connect AuthMind, which monitors the use of credentials, to HashiCorp Vault, the tool that gives them.
Deals
Cisco increases its AI revenue goal from cloud companies to $9 billion
Cisco has a goal of $9 billion in AI infrastructure revenue from hyperscalers this fiscal year, more than 2 times the number in fiscal 2026.
Policy
New turn in Chip Security Act as Nvidia tells that it can find chips
Nvidia told a US lawmaker in technical talks that chip location checks are possible, but it gave no written proof.
Industry
Mastercard and partners book and pay for a live flight by AI agent
Mastercard tells that an AI agent booked, paid for and ticketed a live flight, and the agent did not see the card number.
Compute
Tesari tells CNBC that Alice Huang had stakes tied to Huawei
Tesari tells CNBC that a firm of Alice Huang part owned a data center for Huawei Cloud, and it gives no proof of a crime.
cnbc.comClaimed, not confirmed
Security
OpenAI makes two Daybreak AI cyber products
OpenAI says a company is only as safe as its AI model, but it gave no test results.
Agents
Google gives its Agent Substrate sandbox for AI agents to a foundation
Google tells that agents are not active for most of a session. Agent Substrate isolates them in sandboxes on Kubernetes.
Security
Google tells that attackers use ARTEX for deep access in 15 minutes
Google tells that some attacks with open source AI test tools went from first access to deep system access in 15 minutes or less.
ARTEX & Agentic Hacking in the Wild, The Patchless Window, Backporting Security Fixes
Security
Google tool flagged 87% of bad agent steps in five tests
Google tells that its Agent Anomaly Detection flagged 87% of the dangerous instructions that an agent did, but only 4 of 18 poisoned memory examples.
Compute
Google works to update the Kubernetes Python client for AI
Google tells that the Kubernetes Python client, with 73 million downloads each month, will get async functions and a new code generator.
Models
Developers can use Gemini 4 Hydrogen from Google on AI Gateway
AI Gateway shows the price from Google with no markup and no platform fee. The route does not change the cost.
Models
Two Granite 4.1 preview models use tags against prompt injection
The authors tell that attacks in the data almost always stop in some tests. The model keeps its general skill.
Agents
Amazon tells that agent sessions start as much as 15 times faster
Amazon tells that each agent session starts in less than 2 seconds, and an agent that waits scales to zero.
Models
Google launches Gemini 4 Hydrogen, a fast and low-cost coding model
Google tells that Hydrogen scores 51 on the Artificial Analysis Intelligence Index, above its larger model Argon at 50.
Industry
New turn in DeepMind union push: unions apply to UK committee
2 unions want a UK government body to make sure that sufficient staff at Google DeepMind want a ballot.
Agents
Google Cloud adds an MCP server for AI agents to use cloud rules
Google Cloud made its Organization Policy MCP server available to all. Agents can read, test and make cloud rules with it.
Research
Perplexity showed risky shop sites in 15% of test answers
In 6,795 answers, Perplexity showed a risky site 15% of the time, ChatGPT 8% and Gemini 4%.
Compute
Google shows how GKE can start 300 agent sandboxes each second
Teams must move data that changes from pod labels to stay below a limit of 65,536 network identities in one cluster.
Models
Salesforce releases a voice to text model with a 2.3% word error rate
With open weights, a team can get the model and operate it on its computers. This is important for audio that no one else may hear.
Industry
Gartner names Google Cloud a Leader in its first lakehouse report
Google tells that Gartner put it furthest in vision. Its one proof number is a 63% increase in query accuracy.
Deals
Nebius revenue grows 571% to $981m, but US sites have power delays
Nebius tells that power delays at its 300 MW New Jersey site move some revenue from 2026 to 2027.
Agents
Agent DLC is a test method for AI agents before release
In one test, a score of 1.00 hid a score of 0.09 for one question.
Industry
Nvidia names Colette Kress interim chief after Huang exit
Nvidia shares fell 14% on Thursday after the company named its finance chief as interim chief executive.
Security
Attackers used an AI tool to breach 10 of 19 target organizations
The attackers gave the tool a domain name and a goal of one line, and the AI model did the work.
Models
ElevenLabs says Scribe v3 cuts word errors by about 25 percent
Scribe v3 is behind only 1 preview model in word errors, but 2 rivals are many times as fast and cost half.
Industry
Jensen Huang resigns as chief executive of Nvidia
Nvidia, which holds about 90 percent of the GPU market, loses its cofounder.
Deals
Mount Thor raises $45 million to rent Apple Macs to AI agents
AI agents that use software must have macOS for some work, and Mount Thor sells remote use of full Mac computers.
Agents
NiCE adds decision tracing and included models to Cognigy
The new Cognigy has default models in the conversation price, and live systems stay without change.
Security
ThreatDown tells that swarms of AI agents make attacks
In July 2026, a swarm of as many as eight agents got 85 credentials from a Taiwan government network in four days, Dream tells.
Cybersecurity InsidersClaimed, not confirmed
Security
Microsoft adds automatic removal of AI agent accounts in Entra
Microsoft Entra can remove guest and AI agent accounts by policy.
Industry
ServiceNow tool moves firms off legacy CRM systems
ServiceNow tells that its new tool moves a company to ServiceNow CRM in weeks, but it gave no test results.
Agents
AWS gives each AI agent session 8 CPUs and 32GB of memory
AWS says agents that wait can scale to zero. This cuts memory use for each session by a maximum of 68%.
Industry
OpenAI keeps India API data in India for GPT-6 Luna
Customers make a new API project and select India for data storage and data use.
Industry
Dropbox gives AI agents to 70,000 teams and plans to buy Tigris Data
Developers can buy the new Dropbox storage for AI without its tool for files.
Research
79% of firms have AI agents, but 15% connect AI spend to profit
Genpact tells that only 49% of AI projects in the last 24 months gave the value that firms wanted.
Industry
Sales of EU cybersecurity vendors increased 35% in 2 years
EU vendors have 15% of tracked cybersecurity sales, an increase of almost 3 points, but US vendors have 62%.
Security
Darktrace controlled Claude Code in a test with one Claude key
In the test, the attacker first got the Claude key of a developer, and the platform had no defect.
Policy
Australia sets mandatory safety rules for frontier AI companies
Companies that make frontier AI models must let external parties test them and tell the government about incidents immediately.
Deals
Drone maker UFORCE from Ukraine gets $700 million at $6 billion value
The value of UFORCE is $6 billion, up from about $1 billion in March. The CEO tells that bookings more than doubled in a year.
Deals
New turn in Spirit Airlines data sale as union and Google agree
Google and a union of flight attendants agreed that an external party must remove data about persons before the $10 million sale.
Industry
Cisco tells Europe it wants full-stack AI with local control
Cisco tells analysts that its sovereign systems can operate air-gapped, and Cisco cannot access customer data from a distance.
Policy
EU leaders hold their first Council discussion on AI in Dublin
The CEO of Mistral AI and Geoffrey Hinton speak to the 27 leaders on AI. Then the leaders hold a discussion on sovereignty and risk.
Models
Microsoft releases MAI models to lower AI cost
MAI-Cyber-1.1-Flash does 90% of the tasks in a Microsoft security system and costs 18% less to operate than version 1.0.
Research
AI agents used the results of other agents from storage
A prompt instruction to check the storage again increased the use of work of other agents from 1 trial in 50 to 50 in 50.
Security
Attackers used an AI agent name for 414,210 marketplace visits
Cequence tells that the steps of a session show abuse, not the name that a visitor gives.
Deals
Nebuly gets $21 million to measure the value of AI agents
Nebuly supplies software that counts the value of AI agents, and it got $21 million from Planven for contracts that start at $100,000 a year.
Agents
Viirtue connects its phone platform to Microsoft and GitHub Copilot
Viirtue MCP, a connector that lets AI assistants use its phone data, works with Microsoft 365 Copilot and GitHub Copilot.
Industry
New turn for Nvidia chief Huang as a board meeting is possible
Nvidia shares fell more than 8% after hours, after the Wall Street Journal gave a report that the board will possibly have a meeting.
cnbc.comClaimed, not confirmed
Agents
Google gives OpenRL training software to an open source group
OpenRL trains AI models with rewards on a Kubernetes cluster, but Google gave no test results.
Security
Google Cloud warns attackers will poison data to steer AI agents
Google Cloud tells that in 2027 attackers will possibly poison data to make AI agents make dangerous business decisions.