Agents
AI helped find a Java defect in about one hour
Oracle repaired the defect, rated 9.2 out of 10, on the day of the report. No attack is known.
What matters in AI.
SubscribeNews of the day
58 stories filed under this day.
Agents
Oracle repaired the defect, rated 9.2 out of 10, on the day of the report. No attack is known.
Agents
Microsoft tells that Azure HorizonDB, a database that operates AI models from SQL queries, is available to all customers.
Industry
The Nvidia board told its earnings call that it does not intend to make Jensen Huang chief executive again.
Deals
Anthropic will be the first top AI lab to go public, and its document for buyers names risks to humanity.
Models
The Foundry router selects one of 44 models for each prompt. In some projects it is on from the start.
Deals
MongoDB plans to operate the Firebolt engine with its database, with no more pipeline for customers to make.
Models
OpenAI tells that Astra gets 5.5 points more than Claude Fable 5.2 on one test, at 5% of the cost.
Agents
As of 15 November, maintainers have fixes for 55 of the 99 private reports, and 14 fixes are in a release.
Models
GitHub tells that Fusion Max has 5.4 percentage points more than two other models, at 57% to 88% less cost.
Industry
Google will keep chats with history off for about two months, but rivals keep them for a maximum of 30 days.
Security
Microsoft tells that it sends fixes for more than 1,100 vulnerabilities in October, more than in all of 2025, with its AI scanner.
Security
Hugging Face does not check who is the owner of an account, and one web domain is sufficient to get control of Kokoro.
bleepingcomputer.comClaimed, not confirmed
Models
A decision model is a fast LLM that writes no text and selects from alternatives that are set before, and it is new.
Security
Microsoft tells that an agent must not be the only proof of what it did, and that other systems must give proof too.
Industry
Microsoft tells that agents must have data in one system and the context of that data: what it means, who can see it and if it is correct.
Models
Microsoft tells that its MAI Code 2 model gives 98% of the performance of GPT-6.1 Sol at 40% of the price, in its own tests.
Agents
Microsoft tells that the rules and definitions of a business are more important to agents than the model that they use.
Industry
The Beta is for test work only, and approval for sensitive government data is expected at the start of 2027.
Industry
Databricks says that two catalogs can read the same data copy at this time. The tags move with the data.
Models
Google says Hydrogen controls a real robot arm in eight tabletop tasks and scores 58, almost the same as larger Argon at a lower cost.
Agents
Google says nearly 500 of its customers each processed more than one trillion tokens in the last year.
Agents
Gemini can now book appointments and pay with a spending limit that you set, but Google gave no test results.
Research
Claude Opus 5.5 marked an Iranian school as a target in 72% of trials framed as a game.
Deals
LanceDB tells that customer data stays where it is, with no copy before training starts.
Research
In 27,280 agent runs, Codex CLI was in position 1 by cost for each task with GLM-5.3 and in position 16 with Kimi-K3.
Industry
CNBC reports that the board does not think about Huang as CEO again, and the board will find a new CEO.
Security
The attackers used correct credentials, and no alarm of Uber went off.
Agents
Only 18% of organizations have standard governance for agents, but 49% of developers tell that agents are in production.
Security
One approved sign-in code gave Microsoft 4 methods to misuse a Copilot CLI account in a lab test.
Security
Plaid says its tool found 2.3 times more agent sessions than other tools, with no increase in incorrect flags.
Models
The product lead of its AI platform says that its model work will mostly be post-training of existing models, not new large models from scratch.
Security
A business can let in, limit or stop the AI agent of a customer. Each agent action goes back to a person with a checked identity.
Industry
Pierre Azoulay of MIT Sloan says a platform must set agent access, agent identity and liability before launch. Changes after launch are costly and risky.
Research
The UK Internet Watch Foundation reviewed 6,310 criminal AI images in the first half of 2026. That is 40 percent more than in all of 2025.
Security
Bitdefender Labs found 22 bad GitHub links that make AI coding tools operate code that is not shown.
Security
Six causes can stop the AI agent of Google, also when a site lets it in.
Security
Linx tells that customers control more than 25,000 AI agent identities, and its tools let or stop each agent step.
Deals
Model ML tells that its yearly recurring revenue grew twelvefold, but it gave no agent test results.
Policy
The Irish privacy regulator also speaks with one more firm, with no name given, about AI training on data that agents collected.
Models
The model has 7 billion parameters and operates on a standard CPU at 50 tokens for each second.
Agents
The assistant cannot move funds, and the trader must push Confirm for each trade.
Security
The CEO of Hugging Face, Delangue, says the engineers could examine the July attack only with an open model that a person can download.
Industry
Shoppers will pay for Westwing items in the Zalando app, and Zalando has not said which fee it will get.
Security
The attack that it fights looks safe at first, and usual defenses do not find it in more than 80% of tests by the authors.
Benchmarks
The same tool also gives an incorrect alarm on 70 of 192 safe replies.
Research
A student agent got the habit of its teacher to remove files, although the data had no deletion keywords.
Benchmarks
A judge with no view of culture gives a low risk score to dangerous prompts of one culture.
Research
Six of ten tested Shopify stores show default text with no free delivery limit, return window or bundle exclusion.
Security
After 16 public language model updates, the usual precision score moved by 0.0058 at most and hid the change.
Research
Six models learned less in a chain of eight agents with 300 character notes. They matched one agent only after about four lives.
Research
The models put children above age 13 in 2.5% to 42.5% of audio answers. No warning showed the faults.
Models
It got 93.0% on the test of the authors, against 91.1% for the best of five general models.
Industry
A local partner will operate the cloud, and customers can hold their keys external to Google.
Policy
Before this rule, US regulators had small control of chip access through cloud services in other countries.
Models
Gemini for Google Cloud puts AI assistants in coding, cloud operations and protection tools, but the primary parts are in private preview.
Policy
The US Commerce and Justice departments tell that their Megaspeed export investigation started in 2024, before Nvidia put the firm on its no sell list.
Deals
Cyber seed deals decreased for two quarters in a row, but much of the money went to AI security.
Research
IBM says the top group of leaders get business results from 63% of AI projects, against 43% for all other leaders.