Deals
Cribl buys Tero to cut waste data for engineers and AI agents
Cribl, which collects and routes machine data, bought Tero to decide which of that data is worth keeping.
News of the day
10 November 2026
46 stories filed under this day.
Deals
Mercor raises $750 million at a $20 billion value to test AI agents
Mercor grew with expert work for AI labs. It wants to sell custom tests of AI agents to companies in all industries.
Security
Microsoft shows how to backdoor an AI model with no training
Edits to fewer than 10 rarely used neurons can make an agent run attacker code on a hidden trigger. The model continues to pass standard tests.
Security
New report puts Railway breach at as many as 100,000 workspaces
Railway says the attackers made about $740,000 of AI charges with stolen secrets. Railway will pay what AI providers do not refund.
Agents
Microsoft lets AI agents read Azure data after one sign-in
Each agent step uses the access of the user, and the agent cannot get more than that user can see.
Deals
Founder says OpenAI bought AI travel agent Odessia
OpenAI did not say the deal is correct and has announced no travel product.
Agents
Snowflake Hybrid Tables are as much as 8x faster by default
Snowflake says one small compute unit does the work of eight for some workloads, which can cut cost by more than 80%.
Security
An attacker used an SGLang fault to get data from 18 NEAR AI agents
The confidential mode of NEAR AI did not stop the attacker, because the fault was in software in that mode.
Industry
GoodRx adds drug prices to Google Gemini for healthcare
Care teams can get brand and generic drug prices from Gemini by natural language, in the AI tool that they use.
Compute
Nvidia chief admits spy affair, and AI chip supply is at risk
Companies linked to the suspected spy reportedly enabled $7 billion or more of chip sales to China.
theverge.comClaimed, not confirmed
Security
Proofpoint shows how one approval can put a bad AI agent in a firm
A victim who accepts one blueprint on a Microsoft page can give an attacker a maximum of 250 agent identities in the firm.
Security
The Linux Foundation starts ANS, a name system for AI agents
ANS shows who an AI agent is, but it does not tell the tasks that the agent can do.
Models
Amazon releases Mitra-v2, an open model that predicts from tables
Mitra-v2 had a score above Mitra-v1 on 32 of 33 datasets and is first in single models on TabArena.
Models
Google makes Gemini 4 Carbon open to all paying developers
Google tells that Carbon is first on 3 benchmarks at more than 2 times lower cost than the next external model.
Policy
New turn as EU watchdog opens inquiry into AI adviser Snabe
An EU watchdog opened an inquiry into the selection of Snabe as AI adviser because of conflict of interest concerns.
Security
NVIDIA gives its agent safety tool OpenShell to an open source group
OpenShell sets rules for the steps of an AI agent, and more than 100 organizations use it.
Models
Microsoft releases live speech to text for 61 locales
Final text comes in about 100 milliseconds, and users of this tool make no code change.
Research
Anthropic Fable 5.2 completes 41.3% of freelance tasks
The top agent completed 2.5 percent of the same tasks one year before, and Fable 5.2 completes more than 16 times as many.
Security
Datadog tells that cloud faults are four times more than in May
The median cloud workload had 4 high or critical faults in May and 16 in November, and Datadog links this to frontier AI.
Deals
PuppyGraph gets $10M for AI agents to query connected data
The company tells that PuppyGraph 1.0 completes a query of 10 links in 2.26 seconds, with no copy of the data.
Deals
Bay Security gets $50 million to stop dangerous AI agent steps
The Israeli startup tells that its software stops a dangerous AI agent task before it occurs, with no sensor on each device.
Policy
Eight countries and top AI labs agree to model shutdown controls
The Paris text makes labs agree to shutdown controls for their models, but it sets no deadline and no sanction.
Security
Hugging Face policy chief tells that attacks can close open AI
Irene Solaiman tells the Equity podcast that open platforms are targets because they have open doors and much data.
Industry
Microsoft and IBM launch Imagine Enterprise agent platform
Imagine Enterprise gives companies more than 6,000 made AI agents from IBM on Microsoft cloud, with engineers from the two firms.
Industry
Merz wants Google as partner for a German intelligence cloud
Merz told that Germany must control its technology, and that a US provider can give Europe sovereignty.
Industry
Five UK banks do not let AI agents use customer accounts
HSBC, Nationwide, NatWest, Barclays and Lloyds made the bans before agents tried to use accounts.
Models
boost.ai says its new Presis model has 2.5 times the speed
The company says Presis cuts token costs by 33% on average, but all numbers come from its tests.
Security
80% of banks had attacks from AI agents, a survey finds
BioCatch tells banks to find out if a person or an AI agent controls a session before money moves.
Benchmarks
A new checker cuts the score of one agent from 63.4% to 5.6%
A new test looks at the tools that agents used, and not only at the last answer. It shows that scores can be much too high.
Security
AI agents make new attacks that get 100% on a strong model
Agents that write code made new attacks on AI models, and one attack got 100% success on a model made to stop attacks.
Security
New turn in Fireworks AI breach as Notion finds exposed AI chats
Less than 1% of Notion workspaces are possibly affected, and Notion tells owners to change each password or key that staff pasted into Notion AI.
Security
A monitor external to the model cut AI agent attack success to 1.3%
In a simulation, a continuous attacker drove each prompt defence to 100% success, but a monitor external to the model held dangerous tasks at 0%.
Research
ArmedAgent test finds safety risks when coding agents move robot arms
The test has 36 tasks in 5 scenarios, and the authors report that current systems show substantial safety risks.
Research
Sycophancy in AI models starts in the first training steps
In all 9 tests, models after the first training step had 85% or more of the last sycophancy score.
Security
Claude Fable 5.2 sends more agent tasks to a fallback model
In one test, harm was done in 4.09% of attacks on Fable 5.2 agents, against 1.02% for Fable 5.1.
Industry
ClickHouse buys Xata to give AI agents instant database copies
Xata tells that it makes a copy of a database of many terabytes in about 3 seconds, and ClickHouse gets that skill.
Security
Cloudsmith adds a registry that controls AI agent skills
Cloudsmith uses each skill release as a package with a version and a checksum, and a team can stop one skill or one version.
Security
New turn in Fireworks AI breach after cloud provider hack
Attackers got access to Fireworks development machines on 24 October through a cloud provider and took billing data and some parts of model weights.
Research
GPT-4o does not follow 70.7% of the hard rules in its Model Spec
The same test found that the two Claude models do not follow 0.0% and 9.1% of the hard rules in their specifications.
Research
A small scout model decreases routing cost by 9 to 33%
A small frozen model can predict if a large model will do a task correctly, and how many tokens it will write.
Security
Attack on shared AI sensing breaks 2 defenses with near 100% success
The authors also made STTrust, a defense that finds more than 90% of this type of attack.
Compute
Fused-Softmax gives attention 2.1 to 3.2 times the speed on AMD GPUs
The method works without training, and the error that it adds is less than the error that low precision numbers cause.
Agents
Google DeepMind AI proves important Signal guarantees in 4 weeks
It found no new flaws, but it correctly found known flaws in Signal versions from before their repair.
Policy
Australia sets 6 mandatory rules for AI companies
The government tells companies that "significant penalties" wait for those that break the rules, but it gave no number.
Compute
Google Cloud shows six open AI tools for Kubernetes
Google tells that its llm-d 1.0 router made cache hits two times more, from 35% to 70%.
Agents
AWS adds a web search API for AI agents at $7 for 1,000 queries
The price is $7 for 1,000 queries, and AWS gave no test results for the quality of the results.