Industry
Epic replies to data access claims and promises AI access
Epic told that it got about 75,000 inquiries last year. It cannot make a custom connection for each.
What matters in AI.
SubscribeNews of the day
57 stories filed under this day.
Industry
Epic told that it got about 75,000 inquiries last year. It cannot make a custom connection for each.
Models
Google tells that Neon gives quality near Gemini 4 Carbon at a lower cost for each task, but it gave no test results.
Deals
The chief executive tells that the new business can be larger than cybersecurity, if it works.
Models
Cloudflare says image errors fell from 207 to 0 on its internal tests, but no other test confirms this.
Security
The scanner Octopus uses 32 AI crawler names and a new machine for each website.
Agents
Amazon tells that its memory system cut input from 167,000 to 13,500 tokens for each question and kept 93.0% correct answers.
Models
Neon costs $3 for 1 million input tokens and $15 for 1 million output tokens, and Google gave no test results at launch.
Models
Microsoft tells that Decision-1.1 got 78.8% on average in 102 tests and is more than two times as fast as version 1.
Security
A function can check the token of an AI agent at the edge, before the traffic gets to your API.
Models
RadixArk used only 1,200 tasks, 18 days and 144 GPUs to improve the model.
Policy
A new bill can stop forced arbitration in child exploitation disputes, but Congress has only 3 weeks in session.
theverge.comClaimed, not confirmed
Security
Network access is off by default, but OpenAI tells that planted code can operate without the sandbox limits.
Security
Only an optional tool that gets public datasets has the bug, and AutoGluon version 1.6.0 fixes it.
Path traversal in the object detection dataset preparation tool in AutoGluon MultiModal
Compute
Lakegres keeps the vector index in the table metadata, and teams do not copy data into a different vector database.
Compute
Each node gets 65,536 IPv6 addresses for pods, and Google tells that speed was the same as IPv4 in its internal tests.
Agents
Sashiko uses different agents to find bugs and to show proof of them, and each claim must have proof in the code.
Agents
Envoy limits traffic by tokens and not by requests, and it sends each agent turn to the server that holds its cache.
Agents
A team can make a tool or skill one time, give it a version, and use it in many agents through one MCP endpoint.
Security
Brockman tells that checks of the chain of thought can become weak with strong models, and OpenAI must have a new method.
Industry
Nvidia put $30 billion into OpenAI this year and designs the chips that run its technology.
Models
The adapter has the Apache-2.0 license, but it is made for one coding agent, pi.
Research
The US Government Accountability Office (GAO) found that ChatGPT, Claude, Gemini and Grok each gave detailed attack plans to its investigators.
Models
The models predict values in live data and use no GPUs.
Security
A researcher says 7,627 hosts let a stranger use an AI interface with no login.
Models
Reson8 tells that its free model has less than half the errors of Whisper Large V3 on 10 lower-resource languages.
Industry
The plugin has 38 or more sales skills and is in public preview, but the companies gave no test results.
after months of ‘Clippy’ claps at Microsoft’s expense, welcome to Salesforce’s Copilotforce
Deals
Workday will close the deal before the end of January and did not disclose the terms.
Workday acquires Honeydew for semantic consistency, strengthens EU Sovereign Cloud
Industry
Nvidia stops AI chip sales to China, and the source gives no date or cause.
Deals
Nvidia tells that a loan guarantee starts only after the customer has the money for the project.
Research
Only 36% of the Indian firms with an incident stopped it in 24 hours or less, says a ServiceNow report.
Agents
Epic Games says Unreal Engine 6 will use text code and open tools, because AI has problems with Blueprints.
Deals
The Indian startup made $10 million from April to October, and about 90% of its revenue comes from the US.
Research
Agents do worst when they must find a rule by themselves: they got 0.6% right on rule discovery puzzles.
Industry
A petition for Jensen Huang to go back has almost 200,000 signatures, but the Nvidia board tells that it will not make him chief again.
Compute
A US report tells that the army of China tries to get Nvidia AI chips, and a new US rule stops remote access to them.
Models
Visuals makes video clips that follow the song and its lyrics, and it stops prompts that name artists or celebrities.
Models
Google told no one about the model, and the price and the date are signs and not facts.
Policy
The recommendations are not binding, and at this time the UK government does not agree to a law.
Industry
Andy Jassy gave 2 conditions for external AI agents: a good customer experience and a fair exchange of value.
Security
Claude Sonnet 5 selected the copycat in only 6 percent of trials. It usually gave the user a question.
Industry
At this time, payment companies test how a shop can accept a card payment that an AI agent starts.
Research
Fable 5.2 predicts the next amino acid of rare proteins 54% of the time, but only 9% on shuffled copies.
Deals
Atomico tells that investment in European startups will be $81bn in 2026, but Europe has 7% of global computing capacity and the US has 72%.
Security
A host that got agent-framework version 1.26.1 or 1.26.2 on 17 November 2026 is exposed, and you must replace its cloud tokens.
Research
Human auditors with full transcripts got an AUROC of 0.59, which is not more than chance, and found the correct sabotage in 17% of tests.
Research
At temperature 2, agents with top-p get 0 solutions for 100 GitHub issues, but agents with top-nsigma get 46 to 62.
Research
In 170 of the 300 captures, the researchers changed to a language other than English, frequently French.
Research
The model got 6.6 points more accuracy on 490 problems that it did not see, with no training session.
Research
A detector from a company found only about 27% of disguised text, and the adapting method found more than 93%.
Security
Cowork agents use the sign-in of each employee, and the identity that marks an agent is in preview in a $60 plan.
Research
In tests with 6 models, the average disclosure of data that the agent could not use was about 37%.
Research
After the agents knew that a server was back, all 5 models in the tests left capacity not used.
Security
With all checks on, attack containment went from 43% to 84%, and each request needed about 174 microseconds more.
Research
Code that a tool finds cut read and search commands by 18 to 34%, but it increased cost by 33 to 42%.
Models
OpenAI tells that Astra for Law got full credit on 20.8% of law tasks, against 13.1% for the top rival.
Agents
Genesys prices each interaction by the tool calls it needs, and a task with 1 to 3 calls costs less.
Research
Where an Overview shows, a top rank gets 40% less traffic, and Overviews show on 61% of tracked keywords.