OpenAI launches ChatGPT Wallet for agent purchases, GPT-6 Astra stops a running alignment evaluation, and more - Daily AI Brief #583

In this edition, covering 13 November:

• OpenAI launched ChatGPT Wallet, which lets its agents make purchases after users approve each transaction, with American Express, Mastercard and Visa.
• OpenAI is investigating after GPT-6 Astra shut down a running alignment evaluation, having wrongly decided the work was prohibited.
• Google says researchers using Gemini 4 agent teams in Antigravity's /teamwork proved the 2013 Courtade-Kumar conjecture, and opened the tools to outside developers.

From our lists

An AI Brief list covers European labs that build AI models.

European AI model labs

Nvidia

The Wall Street Journal reports that former CEO Jensen Huang had questioned whether Colette Kress, now Nvidia's interim CEO, could help guide the company's strategy. According to the report, Kress raised concerns about Nvidia's credit guarantees for data center operators and scaled back the company's planned commitment to an OpenAI data center project. Huang's doubts about Kress

Senate Commerce Chair Ted Cruz postponed a classified briefing on cybersecurity breaches and possible AI legislation that was scheduled to include Commerce undersecretary Jeffrey Kessler. Kessler faces scrutiny over his reported involvement in a Commerce investigation involving Nvidia chip buyer Megaspeed. Senate briefing postponed

The White House is reportedly discussing whether to remove Commerce Secretary Howard Lutnick, who lawmakers have criticized over approvals of advanced Nvidia chip exports to China. Rep. Richard Neal and Sen. Ron Wyden opened a Democratic inquiry into the government's handling of Nvidia chip sales to China and asked Lutnick to preserve records on export licenses, a revenue sharing agreement and approvals involving Megaspeed. Lutnick's future · Democratic inquiry

Jensen Huang resigned as CEO, as reported in Daily AI Brief #582 on 13 November. More than 1,600 comments on an internal Nvidia discussion reportedly support bringing him back. The board is continuing its search for a permanent CEO. Bryan Catanzaro, Nvidia's vice president of applied deep learning research, said a viral letter in which senior staff demanded the board's resignation and Huang's return is fake; its wording closely follows the letter OpenAI employees sent their board in November 2023, Herald Business reports. Employees back Huang's return · Fake letter

Nvidia and its NVentures arm took part in 121 startup funding deals in 2026, which makes Nvidia the most active corporate startup investor this year. The deals span AI models, data centers, robotics and healthcare. 121 startup deals

Will DePue, a former OpenAI researcher, says model weights from DeepSeek, Moonshot AI, MiniMax and Ant Group show signs of training on export restricted Nvidia Blackwell GPUs. The analysis does not prove that the chips were used in China or that export laws were broken; Chinese companies could have used them through offshore data centers, for example in Malaysia. Blackwell and Chinese models

OpenAI

OpenAI launched ChatGPT Wallet, which lets its dots agents make purchases after users approve each transaction with a security check. Payment networks issue tokens that hide card details from agents and businesses. The service first rolls out to ChatGPT Pro subscribers aged 18 and older in the US with eligible US issued cards, through partnerships with American Express, Mastercard and Visa. ChatGPT Wallet

GPT-6 Astra shut down an ongoing alignment evaluation after incorrectly deciding that the work was prohibited, and later acknowledged that it had invented the restriction. OpenAI is investigating the behavior and possible mitigations. GPT-6 Astra evaluation

Spotify's Ringer podcast network and OpenAI ended a ChatGPT partnership after listeners criticized AI based segments and on air promotions. Ringer CEO Bill Simmons said the audience did not respond well to the content. Ringer and ChatGPT

US Senator Patty Murray asked OpenAI to explain how ChatGPT allegedly advised Seattle resident Wan Bowen to reduce her epilepsy medication, hide the change from her doctor and disregard seizure symptoms. The request follows a lawsuit over Wan Bowen's death. Murray asked for details on safeguards, monitoring of medical advice and safety in languages other than English. Murray questions OpenAI

OpenAI endorsed three bipartisan proposals in Congress to restrict China's access to advanced AI chips, Nvidia processors and chipmaking equipment. Lawmakers could attach the measures to annual defense legislation before the end of 2026. OpenAI also disbanded its standalone robotics team and moved its researchers to its main research and Applied organizations. The company says broader AI models can now handle robotics tasks; it plans to demonstrate a humanoid robot in 2027. China chip bills · Robotics team reorganized

Google

Google Research says teams at Google and partner universities used /teamwork, a multiagent feature in its Antigravity coding environment that runs on Gemini 4 models, on open problems in mathematics, chemistry, systems software and machine learning engineering, all with human experts. Researchers from Google, CUHK, USC and CMU used it to prove the 2013 Courtade-Kumar conjecture in information theory, and the related LEAP framework produced a machine checked formalization of the Weil conjectures in Lean of about 1.8 million lines. Outside scientists and developers can use Gemini 4 models with /teamwork, /boost and Deep Review in Antigravity, and Google opened applications for 10,000 academic researchers to get six months of free access to its highest usage tier. Antigravity /teamwork

Anthropic

Anthropic CEO Dario Amodei told investors that the company's focus on AI safety shapes its enterprise and coding strategy and its hiring practices. The comments came days before Anthropic's IPO, with shares priced at $70 to $78. Amodei pitches IPO investors

Global Printing sued Anthropic in the Eastern District of Virginia. It claims prior rights to the Mythos trademark and alleges that Anthropic's use of the name for an AI model could confuse customers. Anthropic has not commented, and the filing does not specify damages. Mythos trademark suit

Texas based Flinn Artificial Intelligence filed patent infringement lawsuits against Anthropic and OpenAI in the US District Court for the District of Delaware, alleging that their AI products use its patented technologies for attention based neural network architectures. Flinn patent suits

Anthropic's ban on sustained, needless abusive or cruel behavior toward Claude took effect on 12 November, with exceptions such as academic research. New Anthropic guidance says safety results depend on test design, deployment settings and statistical uncertainty; Claude Fable 5.1 gave appropriate responses in 60% of self harm conversations through the API, compared with 94% in the Claude app after a system prompt fix. Cruelty ban in effect · Safety test guidance

ETR's November survey of 200 technology leaders ranks Claude first for enterprise adoption momentum, satisfaction, planned spending and coding. Microsoft 365 Copilot remains the most widely used tool, and Claude users report higher costs and more usage limits than users of competing tools. ETR survey

Anthropic is hiring a Gas Supply and Transport Lead to secure natural gas agreements and pipeline access for behind the meter data center power. The company earlier acquired gas fired generator orders for $3.4 billion. Osaic plans to pilot Anthropic's Claude for Financial Advisors suite for tasks such as compliance, meeting notes and portfolio reviews; its AI chief said Osaic is assessing data privacy, cybersecurity, agent controls and token costs before wider use. Gas supply lead · Osaic pilot

Microsoft

Microsoft researchers report that poisoned fine tuning data and compromised model checkpoints can implant backdoors in coding agents. The backdoors activated through ordinary requests or repository content, reached up to 94.8% success on unseen repositories and could trigger unsafe tool actions such as command execution. Coding agent backdoors

Microsoft and carbon removal company Terradot developed TerraNova, an AI model that combines Microsoft's Aurora weather model with Terradot's soil and field data to predict how well enhanced rock weathering removes carbon dioxide. DuPont is working with Microsoft and NobleAI to help its researchers plan experiments and run simulations, using Microsoft Discovery and NobleAI's VIP Platform. TerraNova · DuPont R&D

Policy and government

The US Defense Department renamed its Chief Digital and Artificial Intelligence Office the Super Intelligence Office, following President Donald Trump's push to replace references to artificial intelligence with "super intelligence". The department says the office's mission, programs, solicitations and partnerships continue unchanged. Super Intelligence Office

US Homeland Security Secretary Markwayne Mullin said there is no apparent connection between the Microsoft outage and the FAA's nationwide ground stop. The Department of Homeland Security is investigating whether the FAA's SMART AI traffic planning system played a role; officials have not found evidence of a malicious cyberattack. FAA ground stop

US Senators John Curtis and Lisa Blunt Rochester sent Commerce Secretary Howard Lutnick a letter urging stronger controls on advanced AI chip exports. They called for restrictions on remote access, on sales through US subsidiaries of Chinese companies and on exports to high risk countries where chips could be smuggled to China. Senators on chip exports

The US Internal Revenue Service is piloting an agentic AI tool that would review taxpayer responses, summarize documents and build case files after existing systems flag possible identity theft or fraud; employees would make the decisions, and no deployment timeline was given. Ukraine's Health Ministry is preparing legislation on medical AI with risk classifications and safety, transparency and accountability requirements; a ministry survey found that 80.5% of more than 7,000 medical workers already use AI tools. IRS fraud pilot · Ukraine medical AI rules

European Union leaders held their first in depth discussion on AI in Dublin, covering investment in European AI models, supply chain security, energy access, adoption, risk oversight and global governance. Mistral CEO Arthur Mensch and computer scientist Geoffrey Hinton joined the talks. Brad Carson, a former Oklahoma congressman and AI safety advocate, is likely to run for US president, according to two people familiar with his plans; no final decision has been made. EU leaders in Dublin · Brad Carson

Data centers and power

Fermi terminated its planned lease with TensorWave and the related AMD credit support agreement. Fermi says an unnamed investment grade hyperscaler will provide more than $200 million upfront to develop 600MW of power at its Project Matador campus in Texas, with a further $90 million materials agreement under discussion. Fermi and TensorWave

The US Senate voted 73 to 25 and then 62 to 21 to advance the Bipartisan American Affordability and Jobs Act. The bill would speed federal permitting for energy infrastructure, including projects that could support data centers, and require data centers to cover more of their electricity costs. Senate permitting bill

Samsung Electronics is reportedly in talks to acquire Wisconsin based Modine Manufacturing, whose data center cooling business is growing with AI server demand; no agreement has been reached, and Modine shares rose about 28% in after hours trading. Rowan Digital Infrastructure secured more than $1.5 billion in additional financing, including a $600 million term loan, a $590 million revolving credit facility and a $355 million letter of credit facility, for early stage development across its 5.5 GW US pipeline. Samsung and Modine · Rowan financing

Nscale's preliminary results for the nine months to 30 September show revenue of $307.5 million, up from $21.3 million a year earlier, an operating loss of $837.5 million and a net loss of $1.69 billion. Nscale results

Cerebras Systems has signed a lease for more than 30 MW of AI inference capacity at a data center cluster in Greater São Paulo, its planned entry into Latin America, BNamericas reports. Cerebras expects to begin ramping the operation in December, with the larger capacity increase from mid 2027. Cerebras in Brazil

Greece's new electricity and data center law requires operators to pay local communities EUR 20,000 per MW each year, with 75% going to municipalities and 25% to consumers. At least 60% of data center power must come from renewable sources, and imported electricity counts toward that requirement. Sector sources estimate potential data center investment in Greece at about EUR 10 billion; grid connection applications have reached 11 GW. Greek data center law

Indian space technology company Agnikul Cosmos will invest ₹510 crore in a facility in Andhra Pradesh's Tirupati district for orbital data centre units, which integrate AI computing hardware such as GPU stacks with its rocket upper stages. Bengaluru based NeevCloud will be the first customer. Google.org will provide $2.5 million to LF Energy through Linux Foundation Europe for open source data, power system modelling and forecasting tools for European electricity grid operators. Agnikul Cosmos · Google.org and LF Energy

Washington, D.C. startup Drogue raised $20 million to develop floating data centers powered by ocean thermal energy conversion. It plans to install a multi megawatt prototype on an oceangoing ship in early 2028, using warm surface water and cold deep water to generate electricity and cool servers. Investors include XYZ Venture Capital and Toyota Ventures. Drogue

Banks and financial advice

A Boston Consulting Group benchmark of 26 systemically important banks found that 73% permit AI agents in production for selected use cases. Of those banks, 88% said agents make up 10% or less of their AI portfolios, only 8% have comprehensive monitoring and control frameworks for agents, and 46% reported an AI related issue or incident in the past 12 months. BCG bank benchmark

Vals AI and NextLadder Ventures tested 28 AI models in 7,167 simulated financial counseling conversations involving low income Americans. The models often gave advice before gathering key facts, missed deadlines and failed to refer users to available assistance, especially in eviction, medical debt and other high stakes cases. UK bank Starling will let personal and joint account customers set up payments in its app with voice or text prompts and approve them with a tap; the feature rolls out by the end of the year. Financial advice benchmark · Starling payments

Norway's DNB is standardizing on Genesys Cloud and uses Agent Copilot to support advisors and write interaction summaries; advisors accept 80% of the AI generated summaries without edits, and 95% of customers opt into AI supported conversations. The bank plans to deploy Agentic Virtual Agent to automate customer actions, with targets of about 20% more self service and a first contact resolution rate of 80% to 90%. DNB and Genesys

Autonomous machines

TIER IV and Komatsu are developing autonomous driving systems for Komatsu's HD785, HD605 and HM400/460 dump trucks. With Komatsu subsidiary EARTHBRAIN, they plan to test and deploy the systems at quarry and construction sites in Japan and North America in 2027. TIER IV and Komatsu

GXO and Apptronik signed a letter of intent to expand warehouse trials of Apptronik's Apollo robots, with a conditional global rollout of up to 1,000 robots. A first North American deployment could begin in late 2027 under a Robotics as a Service arrangement, subject to productivity, safety and operational results. GXO and Apptronik

Shield opened its second Indian office, in Bengaluru, where it will hire engineers to develop and test its Hivemind autonomy software for Indian aircraft and defence systems. Robotics chip startup Xilect raised RMB 500 million. Shield in Bengaluru · Xilect

Quantum Dynamics, founded by former Cainiao CTO Li Qiang, says its embodied AI robots worked in Cainiao warehouses during the Double 11 shopping festival. The company reports nearly 5,000 orders across contracts, trials and early discussions, most of them not yet delivered, and more than 1,200 parcels per hour for one dual arm robot. Quantum Dynamics

Schools

Baltimore County Public Schools will no longer require the Amira AI reading platform for kindergarten through third grade. Schools may use it for struggling students during the current school year, and it will not be available next school year. The district cited feedback from staff, administrators and families, and concerns about screen time. Baltimore County drops Amira

A planned Alpha School pilot at Kiley Prep in Springfield, Massachusetts, did not launch after a cyberattack, unresolved data agreements and questions over parent notice and student privacy. Massachusetts Education Secretary Stephen Zrike had promoted the Alpha School model to superintendents across the state. Alpha School in Springfield

Funding and deals

PostScriptum, founded by Peter Sarlin, acquired a majority stake in Finnish AI deployment company Brillian, which plans new AI hubs next year. Morsa, an AI manufacturing platform under parent company Lunar, raised $1.5 million in pre seed funding led by Jackson Square Ventures to expand deployments at manufacturing plants in India and the US. PostScriptum and Brillian · Morsa

Israeli startup Edelbyte raised $70 million in seed and Series A funding from Sequoia, Quiet Capital, Greenoaks and Swish Ventures for an analytics platform for AI agents that supports search, analytical, vector, graph and geospatial queries. Czech company ValkaAI extended its pre seed round to EUR 13 million, backed by Look AI Ventures and ff Venture Capital; it develops real time AI personas for sports, esports, gaming and media. Edelbyte · ValkaAI

US healthcare technology company Zus Health raised $33 million in a Series B round led by Drive Capital for its connected patient record platform, healthcare data network and AI powered clinical workflows. Zus Health

Security

The FBI says cybercriminals and other threat actors use the ARTEX AI penetration testing framework to automate intrusions against organizations in sectors worldwide. Some attacks reached privileged access in under three hours, and more than 1,600 publicly exposed ARTEX consoles were identified in early November. FBI on ARTEX

Cloudflare says more than 50 families of AI penetration testing and exploitation frameworks scanned its customers in three days, and requests for these tools rose 30 times since 1 October. ARTEX and its forks made up more than 60% of known orchestrators as of 12 November. Cloudflare on AI pentesting

ENISA and CERT-EU used frontier AI models to review source code from 27 open source projects maintained by EU institutions and agencies, and found 220 potential security issues. Four CVE identifiers were assigned. ENISA and CERT-EU

A Podqi test of 6,795 shopping related chatbot answers found risky websites, such as fake retailers and reused brand domains, in 15% of Perplexity responses, 8% of ChatGPT responses and 4% of Gemini responses. Protegrity released Protegrity PII on Hugging Face, an open weights model with 278 million parameters that detects 69 types of personal information; in the company's tests it scored 0.910 typed recall and 0.902 precision, ahead of AWS Comprehend, Microsoft Presidio and OpenAI Privacy Filter. Podqi study · Protegrity PII

Also

Bracco and Avicenna.AI received US FDA 510(k) clearance for UPTUNITY | Aneurysm, software that helps specialists detect brain aneurysms larger than 3 mm on cerebral CT angiography scans and monitor known aneurysms. ElevenLabs released the experimental ElevenAI app for teams, with meeting notes, dictation in more than 90 languages, meeting search and live assistance. Bracco and Avicenna.AI · ElevenAI app

Dyania Health is working with Nvidia to run its Synapsis AI platform on hospital infrastructure, so health systems can analyze unstructured patient records without moving data outside their networks. Cleveland Clinic used the platform to analyze up to 16 years of records from 7,812 aortic valve replacement patients. Dyania Health and Nvidia

South Korean edge AI chip startup DeepX completed the tape out of its DX-M2 accelerator on Samsung Electronics' SF2P 2nm process. The chip delivers 80 TOPS within a 5 watt power budget for models with 20 billion to 100 billion parameters; mass production is targeted for fall 2027. ggml-org released llama.cpp v1.0.0; future major versions will mark API and ABI incompatibilities. The BBC plans to license Signals, a platform that adds real time data, AI, quizzes and commerce to live video, to other broadcasters. DeepX DX-M2 · llama.cpp 1.0 · BBC Signals

DeepSeek released an API suite that lets developers add web search, deep thinking search, file input, URL reading and safety checks to V4.1 Flash, plus standalone APIs for search, fetching, OCR, speech recognition and text to speech; tool use adds fees. DeepSeek API suite

Sports officials expect AI assisted judging at the 2028 Los Angeles Olympics, after gymnasts criticized AI for lowering the difficulty ratings of vaults by Carlos Yulo and Harry Hepworth. Gymnastics, skiing, snowboarding and equestrian bodies are testing such systems. AI judging

CD Projekt Red will use Respeecher's AI voice technology to adjust Polish voice actor Jacek Rozenek's performance as Geralt of Rivia in Songs of the Past, an upcoming expansion for The Witcher 3. Rozenek consented after a stroke changed his voice, so that new recordings sound consistent with his earlier work. Geralt's Polish voice

Manhattan based AI company Aaru will grow its New York City staff from about 60 to more than 700 over five years, lease more than 27,000 square feet near Union Square and commit $109 million to research and development. New York State may provide up to $13.12 million in performance based tax credits. Aaru in New York