Deals
AI law firm Crosby makes an offer for the brand of Weil and its data
The partners of Weil voted on 10 November 2026 to dissolve the firm, less than 2 weeks after the offer.
News of the day
13 November 2026
49 stories filed under this day.
Security
The FBI tells that the Artex AI tool speeds up attacks
The FBI tells that some Artex attacks went from first target to privileged access in less than three hours. Copies of the tool stay available.
Agents
Google says Gemini 4 agents solved open math and code problems
Google says AI agents helped prove a conjecture from 2013, and the proof in Lean 4 has 50 million lines of code.
Security
Singapore tests the internet address book to identify AI agents
Four firms showed a Chain of Trust plan in October that checks who controls an AI agent.
Agents
Cloudflare opens the code of Kitesurf, its browser for AI agents
Developers can use the code on a Cloudflare account, but a paid plan is necessary.
Industry
Google AI Studio lets developers get Cloud credits in the tool
Developers can make a profile, connect billing and monitor their credit balance in the one tool.
Models
Gemini 4 Hydrogen is open to five GitHub Copilot plans
GitHub adds Google model Hydrogen to Copilot Business and Enterprise as a default model, which an administrator can turn off.
Policy
New turn in Nvidia chip dispute as Treasury names Commerce
Treasury Secretary Scott Bessent told CNBC that Commerce controls the entity list, and that 4 other US bodies wanted Megaspeed on it for months.
cnbc.comClaimed, not confirmed
Models
Protegrity releases an open model that finds 69 types of personal data
Protegrity reports 0.910 recall and 0.902 precision for its model on a test that it made. The score is above AWS Comprehend, Microsoft Presidio and OpenAI Privacy Filter.
Industry
Google cuts Search for schools that stop AI Overviews
Schools with a ban on AI in class must select between AI answers and Search.
Agents
Google adds records of ai agent work to BigQuery and Looker
Owners of agents can record each question, answer and token count in tables of the customer.
Policy
Nonprofits sue five AI companies to stop frontier models
The groups want a San Francisco court to stop new AI models until the companies show control, not damages.
Deals
Google search revenue growth slows to 12% as AI search pays less
Ars Technica reports that an AI Mode query earns less than a usual search. Google gets users faster than ad money.
Policy
Unions win AI limits as 70% of US agencies use generative AI
US government unions win human oversight and protection from AI layoffs in many areas.
Deals
Agentcard gets $8m to let AI agents buy with a user card
Agentcard tells that its tool lets an agent find items, buy and monitor orders at more than 5,000 online shops.
Industry
Anthropic did not sign a letter for open models that 200 firms signed
Nvidia and Microsoft started the letter, and rival firms work as a group against Anthropic.
Industry
OpenAI adds ChatGPT Wallet for its dots agents to buy for users
The dots agents do not see the card number, because the card network gives them a token for each purchase.
Models
ElevenLabs tests ElevenAI, a team app for meeting notes and dictation
ElevenAI is a preview app for teams, and ElevenLabs gave no test results or cost for it.
Research
Meta and OpenAI send 89% of AI agent requests to HUMAN sites
The Muse agent of Meta sent 59% of all agent requests in October, and agent traffic increased 2.6 times in one month.
Agents
Google gives out AQuA, an open source tool for silent agent faults
AQuA finds the cause of a fault in the code of the agent, but it does not make the repair.
Compute
New turn in Nvidia chip oversupply as WSJ report details clash
The Wall Street Journal reports that Colette Kress cut the Nvidia guarantee for OpenAI compute to a maximum of $125 billion.
respawnfirst.comClaimed, not confirmed
Agents
Google Antigravity gives coding work to 2 Gemini 4 models in turn
A planning model and a fast model work in turns on one coding task, but an analyst tells that the documentation gives no record of each step.
Industry
News union fights USA Today Co. on its Palantir deal and AI tools
The NewsGuild-CWA, the largest news union in the US, wants a written contract promise that AI will not replace journalists at USA Today Co.
Research
AI found 99 vulnerabilities in network code in 90 hours
Forescout says the same work by hand needs nine months, and the AI cost only $1,200 to $2,300.
Deals
New turn in OpenAI funding at $15B and $1.25T
CNBC says OpenAI will get $15 billion at a $1.25 trillion valuation, half of its $30 billion goal.
Policy
New turn in Nvidia China chip scandal as Lutnick agrees to testify
The US Senate Banking Committee wants the Commerce Secretary and the Nvidia chief to testify in December about chip sales to China.
cnbc.comClaimed, not confirmed
Research
Firecrawl and Apify get more right answers than Claude web search
On protected sites, Claude web search gave 49% right answers, and Apify gave 98% right answers.
Security
Anthropic investigates sign-in errors on 4 Claude services
The notice gives no cause, no start time and no count of users.
Industry
Zendesk opens its support software to ChatGPT and Claude agents
Zendesk Bridge lets external AI agents work in Zendesk, and Zendesk sets outcome-based pricing.
Security
A new attack makes AI agents send 93.88 times more calls
The attack keeps an AI agent at work after its task is complete, and the provider pays for the additional work.
Security
Edits to fewer than ten neurons can make an AI agent a sleeper
The method, Lazarus, hides a backdoor in model weights with no training, and seven neurons caused the payload in 98.8% or more of the triggered tests.
Research
Incorrect risk warnings made Gemini refuse 86.3% of safe tasks
An attacker can stop an AI agent if the attacker puts an incorrect warning in content that the agent reads.
Security
Schwab says it stops AI agents 10 to 20 times each week
Large banks stop AI agents that use customer accounts because they cannot tell an agent with authority from an attacker with the same login.
Security
A $20 attack through a training service can remove AI model safety
The researchers propose Bait, a defense that probes the reward function to decrease the attack.
Research
A clean history can make an AI agent worse when the error remains
A success receipt on a defective database write cut task success from 45% to 28%.
Research
AI agent memory records remove the sign that a user fact is temporary
In all 11 models tested, memory records change is working at to works at, and not the opposite.
Security
BELLS-O tests 33 AI safety filters for speed and cost
In moderation, specialized filters have 93 to 96% detection, as the best general models do, and operate 5 times faster.
Benchmarks
Authors remove a claim that broadcasts increase agent hallucinations
A judge counted correct reports of a teammate's error as hallucinations, and the new effect is -0.0208, near zero.
Agents
Logos SDK kept all 197 faulted agent sessions in a test
With one fault in each of 200 tasks, the single-process baseline kept none of 197 faulted sessions, but the Logos SDK kept all and wasted no call.
Research
Runtime incentives increase hard-task agent success from 0.60 to 1.00
Text in a prompt that promised a reward had no detectable effect on agents in 240 runs, but a rule in the runtime did.
Research
Teams of 30 small AI models cut a single model's error by only 1.7%
The models make the same error on the same item, and a team of thirty cuts error by a median 1.7%.
Compute
MARS cuts the wait of AI agents on GPUs with small memory
The authors report that MARS cuts mean latency by 1.59 to 7.56 times against the fastest baseline.
Policy
Singapore charges a fifth person in the Nvidia server fraud case
Singapore also barred Alice Huang, a businesswoman in the Nvidia scandal, from re-entry to the country.
Research
IBM finds 94% of firms had a digital sovereignty problem
Only 33% of firms could fully show that they can replace a provider, model or system.
Agents
Brave Search lets AI agents pay $0.005 for each search with x402
An AI agent can make a search with Brave and pay for each request in a stablecoin, with no account and no API key.
Agents
Schwab starts an AI server for trades, first with Claude
Schwab, a US broker with $13.33 trillion in client assets, lets clients make trades through AI, and the client must give approval for each trade.
Models
Nous Research cuts propaganda in a Qwen model from 22 to 1 percent
Nous made the test, and the company gives no external test result.
Industry
JPMorgan stops some AI agent purchases made with virtual cards
JPMorgan puts the virtual cards of AI agents in the class of commercial cards, and this stops some agent purchases.
Industry
Claude Fable 5.2 moves above Gemini 4 Carbon on a top AI index
The source gives no scores for the 2 models.