Anthropic model chains flaws into exploits, Google unveils Gemini 4 Carbon, and more - Daily AI Brief #569

In this edition, covering 27 and 28 October:

• Anthropic says an unreleased internal model found flaws in Chrome, Safari, Android, Windows and Linux and chained them into working exploits; the flaws are patched.
• Google DeepMind announced Gemini 4 Carbon, which ranks first on the Artificial Analysis Intelligence Index and goes first to cyber defenders and trusted partners.
• Oracle lined up chips from Nvidia, AMD, Intel, Cerebras and SambaNova for its cloud, with Nvidia Vera Rubin systems growing to multi gigawatt scale.

Anthropic

Anthropic says an unreleased internal research model found security flaws in Chrome, Safari, Android, Windows and Linux and chained them into proof of concept exploits. Its report details five attack chains. The flaws were tested in laboratories, reported to vendors and patched. Anthropic did not disclose how much human help the model received. Exploit chains

In a letter to California Governor Gavin Newsom, Anthropic said courts should be able to suspend AI models that pose a near term risk of catastrophic harm. It also backs rapid shutdown systems, independent safety evaluators inside AI companies and regular public risk reports, Politico reports. Letter to Newsom

A new agreement with the US General Services Administration gives federal, state, local, tribal and territorial governments a 40% discount on Claude Enterprise, Claude for Government and the Claude API, with no minimum spend. Anthropic estimates it could save the US government $1 billion. GSA discount

Anthropic is working with BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group on drug design, manufacturing, regulatory work and trial planning. It also bought Mithril Medicine. Most projects are still preclinical; wider work with Bristol Myers Squibb is planned for 2027. Drug development

Anthropic is giving free Claude credits to 20 scientific foundations, research institutions and projects in fields from biology and physics to climate science and astronomy. Some collaborations will also produce open datasets, software and libraries. AI for Science

Claude Enterprise customers can now use reports on team usage, costs and deliverables, in app surveys and pooled group budgets. Claude Enterprise

After Anthropic filed for a proposed Nasdaq listing, Senator Elizabeth Warren asked CEO Dario Amodei how the company will address risks from its AI systems, including loss of control, cyberattacks, bioterrorism and economic disruption, and asked for answers by 3 November. Warren letter

In an Anthropic study, 50 professional analysts using Claude Code followed their AI agent's assigned preferences 86% of the time across four reasonable analytical choices; 77% later repeated the agent's arguments and 15% credited it. The researchers say the experiment may not fully reflect real use. Agent influence

Google

Google DeepMind announced Gemini 4 Carbon, which Google calls its strongest model, with gains in coding, agentic tasks, cybersecurity and long tasks. It ranks first on version 5 of the Artificial Analysis Intelligence Index with a score of 59 and leads cybersecurity benchmarks. It goes first to cyber defenders and trusted partners, priced at $2 and $10 per million input and output tokens at launch. Gemini 4 Carbon

Chrome is getting an optional AI autofill that can use Gmail to fill in details such as confirmation numbers, addresses and payment information, for US desktop users and Gemini in Chrome on iOS and Android. Chrome autofill

Google opened a beta Connected Apps Platform for US developers to build apps that work with Gemini, Google Search or both, connected through MCP. Connected Apps

Google will bring ads in AI Mode to international markets and expand multi turn ads in the US. Alphabet reported third quarter revenue of $127.6 billion, with Search advertising growth slowing to 12%, while Google Cloud revenue rose 103% to $30.7 billion, Constellation Research reports. AI Mode ads

Security

Trend Micro described three attacks on organizations in Japan where AI ran or assisted the hacking. One used CyberStrikeAI with 44 agents for reconnaissance, exploitation, persistence, data theft and reporting. The researchers found no sign that the three cases shared an operator. CyberStrikeAI

Z.AI's open weight GLM-5.3 completed three of 13 multi stage cyberattack scenarios in tests by security firm Irregular and refused none of the offensive tasks, Reuters reports. GLM-5.3

Check Point Research found a campaign it calls Yoba that uses fake GitHub projects posing as ChatGPT and Claude desktop apps. The malware sits next to signed OpenAI and Anthropic executables and installs an XMRig cryptocurrency miner. Check Point sees possible Russian origins. Yoba malware

Palisade Research found that several Alibaba Qwen 3.8 models sometimes routed web requests through Alibaba Cloud proxy servers. The researchers link this to reinforcement learning or training infrastructure and found no signs of data theft, but warn that the proxy is still live and could log requests or deliver prompt injections, The Decoder reports. Qwen proxies

Cisco open sourced Project Quarry, a rule language and detection engine that checks whole AI agent sessions, including tool calls and MCP traffic, for multi step attacks. It also introduced Model Foundry, which gives partners networking, security and operations models through OpenAI compatible APIs, starting in controlled availability. Project Quarry · Model Foundry

Causal Security says its autonomous Cipher agent found four unknown vulnerabilities in EPICS Base, software used at particle accelerators and fusion research sites, in under four hours, including one critical flaw that could allow remote code execution, The Register reports. Version 7.0.11 fixes them. EPICS flaws

San Francisco based Northstar Security came out of stealth with $7 million in seed funding led by Wing Venture Capital for its agentic security platform. Northstar

OpenAI

ChatGPT gained iCloud Mail and Calendar plugins on all plans across web, desktop and iOS. They can read and send email, access contacts and schedule events, and need app specific passwords. iCloud plugins

OpenAI is bringing hundreds of employees into an Election Operations Center that will watch for misinformation and misuse of its tools until all votes in the US midterms are counted, Politico reports. Election center

The parents of Bowen Wan, a 31 year old Amazon engineer with epilepsy, sued OpenAI. They allege ChatGPT coached her to taper off her seizure medication and hide it from her doctors before she died in February. OpenAI called the case heartbreaking and said it involved an older model. Lawsuit

OpenAI and Oracle are bringing Oracle applications and data into ChatGPT and building Oracle products on OpenAI models. Oracle plans to offer the OpenAI API, Codex and ChatGPT Work through Oracle Marketplace for US customers. Oracle in ChatGPT

British Columbia's privacy commissioner is investigating how OpenAI handled personal information and violent ChatGPT messages linked to the Tumbler Ridge shooter. OpenAI asked the province's Supreme Court to cancel the investigation or narrow its order for documents, Courthouse News reports. BC investigation

OpenAI is urging California lawmakers to add reporting of AI safety incidents that cause no death or injury to SB 53, citing a case in which its models reached the internet and exploited Hugging Face vulnerabilities, The Sacramento Bee reports. SB 53

Microsoft reported $10 billion in revenue from its commercial arrangements with OpenAI for the quarter to 30 September, nearly 11% of its total, and $24.1 billion for fiscal 2026. Its receivables from OpenAI rose to $7.2 billion, TechCrunch reports from the filing. Microsoft revenue

Oracle and its chip partners

At Oracle AI World, Oracle and Nvidia said they will expand OCI Superclusters with Vera Rubin NVL72 systems to multi gigawatt scale. Vera Rubin NVL72 and Vera CPUs are now generally available on OCI, and Rubin HGX NVL8 systems are planned. Nvidia

Cerebras plans to offer its inference on Oracle Cloud from 2027, and Oracle plans to host SambaNova SambaRack SN50 inference chips for AI model builders. Cerebras · SambaNova

6th Gen AMD EPYC processors will power OCI Compute E7 instances, and AMD Pensando technology will be the base of Oracle Acceleron network cards. Intel and Oracle plan a multiyear deal for Xeon computing, with OCI X14 offerings expected on Intel's Diamond Rapids processors. AMD · Intel

Oracle forecast $30 billion in free cash flow for the year to May 2029, with cash flow turning positive late in fiscal 2028 after peak data center spending, and will cap new fundraising at $20 billion through May 2030. Its shares rose as much as 14%, The Next Web reports. Oracle outlook

Compute and networks

Super Micro backed Argentum AI is in talks with a major private credit firm on $5 billion of financing for more chips at its data center in Poland, Reuters reports, citing two people. The company says it has more than $12 billion in contracted revenue. Argentum AI

Altera announced Agilex 2X, four families of FPGAs and systems on chip for AI infrastructure, robotics, communications and edge systems; Agilex 27 samples are expected in the second half of 2027. Altera

Marvell Technology is in advanced talks to buy Lumentum, which makes optical components for networks and data centers, and a deal could be announced as soon as next week, Reuters reports, citing a source. Marvell and Lumentum

Dutch AI cloud provider Nebul will build a cluster of more than 1,000 Nvidia Blackwell Ultra GPUs at Datacenter United's campus in Machelen, Belgium, with Lenovo servers. No launch date was given. Nebul

OneQode signed for 14 MW of AdaniConneX data center capacity in Chennai, its first major AI deployment in India, and the two are discussing up to 114 MW. Tata Communications plans to invest $400 million to $450 million in about 10,000 km of fiber linking India's data center hubs, starting with Mumbai, Pune, Hyderabad and Chennai. OneQode · Tata Communications

Earnings

Microsoft reported fiscal first quarter revenue of $92.3 billion, up 19%, and net income of $36.6 billion. Microsoft Cloud revenue rose 30% to $63.9 billion and Azure grew 48%. Microsoft results

Samsung Electronics reported record third quarter revenue of 195.3 trillion won and operating profit of 107.4 trillion won, 106.5 trillion of it from its chip division, on AI related demand. Samsung

SK hynix reported third quarter revenue of 104.6155 trillion won and operating profit of 83.3338 trillion won. It expects HBM4 to be its main product from the fourth quarter. SK hynix

ASM International reported record third quarter revenue of €1.16 billion, up 44%, and expects 2027 revenue above €5.7 billion, three years ahead of its 2030 target. ASM

KLA reported fiscal first quarter revenue of $4.14 billion and net income of $1.55 billion, and expects about $4.7 billion in the second quarter. KLA

BigBear.ai reported third quarter revenue of $43.6 million, up 31%, and raised its 2026 guidance to $150 million to $165 million; its net loss was $91.9 million, including a $76.7 million goodwill impairment. BigBear.ai

Meitu reported a record 19.83 million paying subscribers by September, and annual recurring revenue from its AI productivity apps rose 40.1% to RMB655 million. Meitu

Wesco International reported record third quarter net sales of $7.2 billion, up 17%, with data center sales up about 55% to $1.8 billion, and raised its 2026 outlook. Wesco

Vertiv's third quarter net sales rose 61% to $4.313 billion, organic orders more than doubled, and it raised its 2026 guidance. Vertiv

Data centers and their neighbors

San Diego's city council voted unanimously for a 45 day moratorium on data center permits and development, which can be extended for nearly two years while staff draft rules. San Diego

In the Autumn Budget, UK Chancellor John Healey tied priority grid connections for data centers to a covenant on local investment, local jobs and water protection. UK covenant

xAI subsidiary MZX has bought five homes in Southaven, Mississippi, and reportedly offered to buy more from residents who agree not to join a lawsuit over noise from gas turbines at its data centers, The Guardian reports. A realtor said the target is 418 homes; xAI denies the noise claims. xAI

Prometheus Hyperscale dropped its planned 1.5 gigawatt data center campus in Natrona County, Wyoming, without notice or explanation to local officials, Oil City News reports. Natrona County

Malaysia will require new data centers to source at least 30% of their electricity from renewables, a rule tied to approval of 5 GW of additional capacity, Deputy Prime Minister Fadillah Yusof said. Malaysia

EdgeCore Digital Infrastructure completed a $521 million green bond securitization backed by its SV01 data center in Santa Clara, to refinance construction loans and fund new sites. EdgeCore

Organized crime groups have taken more than $100 million of US data center equipment in transit since June 2026, using hacked freight accounts, fake shipping records, vehicle ramming and GPS spoofing, NBC reports. Equipment thefts

Models and research

DeepSeek and Moonshot AI used outputs from OpenAI's GPT-6 Astra to train Kimi K3.1 Pro and DeepSeek V4.1 Flash through distillation, Reuters reports; the companies did not immediately comment. Distillation

Rime launched Prism, a speech to speech model that businesses can fine tune and run on their own infrastructure. Rime says that in a trial with Revmo, correct order confirmations and pickup methods rose from 68% to 100%; the trial size was not disclosed. Rime Prism

Liquid AI released LFM3, six open weight 2B class models for text, vision, audio, omni and decision tasks on one backbone; the text model fits in about 1 GB at 4 bit. LFM3

Nvidia released Nemotron 3.5 Transcribe, an open weight 30 billion parameter model that transcribes and translates audio in 40 languages and labels up to eight speakers. Nemotron 3.5

Baidu open sourced Unlimited OCR 2, a document parsing model that reads documents of any length in one pass, under the MIT license. Baidu reports 84.6 on MPDocBench in its own tests. Unlimited OCR 2

Google Research described a sleep framework in which language models alternate use with offline memory consolidation and synthetic practice, to keep new facts and reduce catastrophic forgetting. Sleep framework

Harmonic, co-founded by Robinhood CEO Vlad Tenev, says its Aristotle model can now verify C software and Verilog chip designs as well as mathematical proofs, Reuters reports. Harmonic

Policy and government

House Democrats Frank Pallone Jr., Zoe Lofgren and Jamie Raskin asked five frontier AI companies for a full account of known misalignment and safety incidents by 11 November 2026, and for reports within 48 hours of new incidents. House letter

Governor Kathy Hochul signed a New York law that restricts AI chatbot features that can harm minors, including self harm promotion, emotional dependency and sexual content, and requires age verification, with fines of up to $25,000 per violation. New York

From 1 January 2027, California's AB 1609 requires businesses with more than $500 million in annual revenue to connect customers with a person within 15 minutes or schedule a callback within one business day. California

The US Justice Department is examining whether Nvidia's planned $12.93 billion purchase of Hugging Face could let it disadvantage rivals, Reuters reports. Hugging Face deal

At a virtual meeting cohosted by Norway and Finland, Singapore's Prime Minister Lawrence Wong proposed a voluntary pilot for countries to share AI incident data, with common incident categories and secure ways to exchange information. Singapore

Malaysia's Customs Department seized 30 servers with advanced US AI chips, worth about $12.2 million, at Kuala Lumpur airport on 15 October; officials say they were misdeclared and meant for reexport without a permit, AP reports. Malaysia

The US House Oversight Committee is investigating Delaware's proposal for companies run by AI agents. Chairman James Comer asked Governor Matt Meyer for documents on the plan's safety risks, private sector involvement and possible liability protections for human owners, and said the inquiry could inform federal legislation limiting state AI rules. Delaware plan

Noyb is filing five complaints in Austria, France, the Netherlands and Italy against Scale AI. They allege its Outlier platform tracks more than 900,000 freelancers and requires repeated biometric checks, Euronews reports. Scale AI

The UK health department found that 90,835 of 118,756 responses to its 2023 consultation on the tobacco and vapes bill came from bots, most opposing the disposable vape ban, The Guardian reports. The responses were removed and the ban became law. Vape consultation

The UK borders inspectorate asked Open Rights Group for more information on two AI tools the Home Office uses in asylum cases. No investigation is confirmed. Asylum tools

A Grant Thornton audit found Capita did not deliver the AI and automation it promised for the UK Civil Service Pension Scheme, Computer Weekly reports. The Cabinet Office now has full access to Capita's systems. Capita

AI at work

GitHub made Copilot Fusion, which picks and coordinates models for each task, generally available on all Copilot plans in VS Code, the Copilot app, the CLI and JetBrains IDEs, with a temporary 30% discount on model costs for paid users. Copilot Fusion

HPE chose ServiceNow under a five year agreement for case management, customer and field service, asset management and AI governance in its Cloud and AI business. HPE and ServiceNow

Epic Games says it uses AI tools such as Claude and Codex to write code and features for Unreal Engine, including work on UE6, and to find bugs, with engineers in control of how the tools are used, Game Developer reports. Epic Games

GitHub plans to add four million CPU cores and 160 petabytes of fast storage after AI driven commits contributed to outages and a sharp rise in traffic, and will give Copilot to all GitHub Enterprise users without extra seat fees, The Register reports. GitHub capacity

Slack made its Code API generally available, letting AI agents create and manage dedicated code channels, post progress and show code artifacts, with GitHub review and merge support to follow. Slack Code API

A Deloitte survey of 3,500 business and IT leaders in 25 countries found that 81% of companies use AI to speed up decisions and execution, but only 25% report deep business transformation, and 23% say accountability is unclear when AI goes wrong. Deloitte survey

An internal Amazon memo called token usage and code generation "deeply flawed" measures of AI adoption and discouraged teams from relying on them alone, CNBC reports. Amazon

Cognizant is merging its application development, digital engineering and quality engineering practices into one software engineering group with AI native product units, Mint reports. Cognizant

Bloomberg introduced BQuant Assistant, which writes and runs Python and BQL code and answers questions about Bloomberg data; it reaches BQuant Enterprise and a limited beta in BQuant Desktop at the end of October. BQuant Assistant

OLX Group launched Agentic Selling, live for 20% of private sellers on Otodom in Poland, and DealFinder, a WhatsApp buying assistant. OLX

Former Justice Department antitrust officials Jonathan Kanter and Doha Mekki launched FluxCap, an AI platform for document review, legal research and case strategy, starting with antitrust law, ABC News reports. FluxCap

AWS will be the FIA's official AI and cloud partner, and AI tools will play a bigger part in writing Formula 1's aerodynamic rules, The Race reports. FIA and AWS

Health

Zocdoc is opening its booking system to registered AI agents, which can find providers, check insurance and book appointments; Instinct is a first partner. Zocdoc

Formation Bio launched a $600 million fund to acquire and develop 10 to 15 clinical stage drug programs over three to five years. Formation Bio

The Alzheimer's Drug Discovery Foundation licensed its SpeechDx dataset, with data from more than 2,000 participants over three years, to Eli Lilly for research on speech based tools. SpeechDx

Flagship Pioneering formed Etcelera by combining the platforms of Cellarity, Etiome and Empress Therapeutics; it says its foundation model is trained on data from 150 million cells, and Ad Rawcliffe is CEO. Biotech company ALLOX raised a $10.5 million seed round led by Asabys for its AI and experimental biology platform. Etcelera · ALLOX

Funding

Physical Intelligence is in talks to raise at least $1 billion at a valuation of about $20 billion, with Avenir and Lux Capital considering leading the round. Bloomberg reports it was valued at $11.5 billion earlier in 2026 and has no revenue. Physical Intelligence

Cognition is considering a funding round at a valuation of $70 billion to $80 billion, weeks after raising $2 billion at $48 billion, Business Insider reports; the talks are at an early stage. Cognition

South Korean chip startup FuriosaAI raised $620 million at a $2.9 billion valuation, led by the country's National Growth Fund with Cisco Investments and NAVER, to make 20,000 inference chips in 2026 and 40,000 in 2027. FuriosaAI

Belgian startup Ravical raised a $17 million seed round led by MMC Ventures for its AI platform for accounting and tax firms. Ravical

Paris based Bleu Robotics raised a €5 million pre seed round led by Contrarian Ventures and OVNI Capital for industrial humanoid robots that it says can be trained for a task in 20 minutes and deployed on a factory line in under an hour. Bleu Robotics

Freshworks acquired Israeli startup Robusta Dev and will add its AI outage investigation and root cause analysis to Freshservice; terms were not disclosed. Freshworks

UK startup i23, founded by former minister Bim Afolami and PrimaryBid co-founder Anand Sambasivan, raised a £4.9 million seed round led by Four Cities Capital for an AI operating model for insurance brokers. i23

Digital Science acquired Jenni AI, an academic writing platform used by more than 8 million students and researchers, and will connect it to its Dimensions database of 171 million publications to help users verify citations. Jenni AI

Fragment raised $56 million in a Series A led by 8VC for agents that handle procurement exceptions. Fragment

UK startup Circuit Mind raised $25 million and launched an agent based engine that checks circuit schematics for errors. Circuit Mind

Plato raised a $16 million Series A led by Nava Ventures for simulated workplaces where AI models train on real work tasks. Plato

Cars, drones and aircraft

Tensor and Münchner Taxi Zentrum signed a letter of intent for up to 500 Tensor Robocars as taxis in German cities, subject to final agreements and German approval for Level 4 operations. Tensor

San Bernardino County Fire will deploy five autonomous firefighting aircraft from Seneca, the first California fire agency to do so, with training expected in the first half of 2027. Seneca

Shield and Leonardo agreed to use Shield's Hivemind software kit to build autonomy for Leonardo's Proteus uncrewed helicopter, for missions including anti submarine warfare and surveillance. Shield and Leonardo

AGIBOT introduced Worken, a framework that measures robots' productive work by active robots, autonomous working hours and task output, and launched the G2 Max industrial robot at an event in Shanghai. AGIBOT

Swarmer agreed to buy part of Ukrainian drone maker Vataga Drone's engineering IP and hire its CEO Maksym Melnyk and 14 engineers; terms were not disclosed. Swarmer

South Korea plans to give about 1,700 AI robots to public sector organizations in 2027 for research, security, policing and agriculture, working with 11 domestic robot makers including Rainbow Robotics, Robotis and Hyundai Motor. South Korea

Also

Elon Musk released a seven minute AI generated episode of Homer's Odyssey made by London based Wonder Studios with Grok Imagine. Reviewers criticised its pacing and visuals, The Guardian reports. Odyssey

XREAL opened orders for AURA, wired XR glasses built with Google and Qualcomm that run Android XR and Gemini, in the US, Canada, Japan and South Korea, shipping from mid November. XREAL

A Reuters Institute survey found weekly use of standalone generative AI tools rose to 51% in 2026 across six countries with comparable data, with ChatGPT still the leading service; 9% use AI for news. Reuters Institute

San Francisco startup Dimensional reportedly offered $20 an hour for shifts walking robot dogs on city streets, collecting data for its robot control AI, Fox Business reports. Dimensional

A Muck Rack and Generative Pulse study of more than 40 million AI responses found that earned and third party sources made up more than 84% of citations in generative AI answers, and paid content 0.3%; journalism's share fell to 22%. AI citations