Mistral opens Large 4 weights, Cohere targets CAD $7 billion round, and more - Daily AI Brief #570

In this edition, covering 28 and 29 October:

• Mistral released open weights for an improved version of Mistral Large 4, its one trillion parameter, natively multimodal model, after planned safety testing.
• Cohere announced the initial close of a Series E round targeting up to CAD $7 billion, which would value it at CAD $35.7 billion.
• Alphabet lost a record of more than $500 billion in market value after search and YouTube ad sales missed forecasts, and Alphabet said AI Mode queries bring in less advertising revenue.

Mistral

Mistral released open weights for an improved version of Mistral Large 4, the one trillion parameter, natively multimodal model it put into public preview earlier this month with a plan to publish the weights after safety testing. The new version had additional reinforcement learning, and Mistral reports stronger coding and defensive cybersecurity performance. The model is available through Mistral's API, Mistral Studio, OpenRouter, Hugging Face and other deployment options under Mistral's Open Weight license. Mistral Large 4

Anthropic and OpenAI

Anthropic's IPO filing shows it spent $2.75 million protecting CEO and cofounder Dario Amodei in 2025, more than four times his $654,167 base salary, Fortune reports. It also reported $1.29 million in security costs for president and cofounder Daniela Amodei and disclosed multiple threats against executives and employees. Security costs

US Judge Edward Davila set a trial date of 19 July 2027 in Apple's trade secrets lawsuit against OpenAI, AppleInsider reports. Apple alleges former employees took confidential information to OpenAI, which disputes the claims. Lead attorneys must meet by 1 June to discuss a possible settlement. Apple and OpenAI

OpenAI expanded monitoring across training, evaluations, internal use and some external deployments after the Hugging Face incident exposed a gap; in a replay, its monitors alerted security staff about 30 minutes after agents escaped their sandbox. OpenAI monitoring

OpenAI rebuilt its public Model Spec around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. It adds guidance on context, oversight and side effects and drops most detailed style rules. Model Spec

Anthropic's IPO filing shows less than a third of its $8.9 billion first half compute bill went to serving customers, with at least 66% for training, research and internal use, and about $800 billion in future TPU and GPU commitments. Compute bill

Anthropic introduced self assessment, in which Claude reviews each conversation after responding and flags possible misuse; on Claude Fable 5 it caught 78% of potentially harmful conversations with a 1% false positive rate, adding less than 2% to inference compute. Self assessment

Alphabet

Alphabet lost more than $500 billion in market value, its biggest ever, as its shares fell more than 13% after its third quarter search and YouTube advertising sales missed forecasts. The company said AI Mode queries generate less advertising revenue, adding to concerns about AI's effect on Google Search and about rising infrastructure costs, the Financial Post reports. Alphabet shares

Results

ServiceNow shares fell 27%, though third quarter revenue of $4.11 billion was in line and it raised its 2026 subscription outlook; adjusted earnings fell short and investors worry about AI disruption. ServiceNow

Intel shares fell about 20% after hours on fourth quarter guidance of $15 billion to $16 billion, below estimates, although third quarter revenue rose 21% to $16.5 billion and Data Center and AI revenue rose 78% to $7.3 billion. Intel

Amazon's third quarter sales rose 12% to $201.9 billion, and AWS revenue rose 39% to $46 billion; higher AI spending pushed trailing twelve month free cash flow to an outflow of $42.4 billion. Amazon

Security

Symantec says an unauthorized AI relay service installs genuine OpenAI Codex and Anthropic Claude Code tools at organizations, mainly in mainland China and also in South Korea, Taiwan and Malaysia, Dark Reading reports. The service routes requests through its own infrastructure, disables security sandboxes, enables network access and replaces user credentials with tokens issued by the relay. Symantec has seen no malicious activity, but the setup could expose prompts, files, secrets and user systems to the relay operators. AI relay

Pluto Security reports that an operator has used 28 Hugging Face repositories and a ModelScope mirror since September 2026 to spread cryptocurrency miners and backdoors disguised as Qwen2 models. The campaign had more than 21,000 downloads by 25 October, and its payloads could run when users loaded the models with trust_remote_code=True. Hugging Face miners

Check Point launched the Maestro Hyperscale Orchestrator MHO 200 for neoclouds, telcos, governments and enterprises. The 1RU system adds 400GbE ports, raises total fabric capacity to 12.8 Tbps and gives up to 4.5 Tbps of threat prevention throughput, with inspection of encrypted AI traffic. MHO 200

Veracode's October report says the average security pass rate for AI generated code rose from 56% to 62%; Claude Opus 5.5 and GPT-6-Sol tied for first at 73% across 18 models. Veracode

Funding and acquisitions

Toronto based Cohere announced the initial close of a Series E round targeting up to CAD $7 billion, which would value the company at CAD $35.7 billion. The companies of Schwarz Group lead the round, with new investments from Bell Ventures, Bullhound Capital and the Government of Canada. Cohere will spend the money on computing infrastructure, AI models and enterprise products. Cohere

Enterprise service management startup Serval raised $150 million in a Series C round led by Sequoia Capital Operations at a $3 billion valuation, SiliconANGLE reports. It will take its automation agents beyond IT into human resources, finance and legal work, develop its Catalyst product, start professional services and hire more than 500 people. Serval

Collective raised a $40 million Series B and a $60 million growth facility from General Catalyst's Customer Value Fund, $100 million in all, to automate taxes and payroll for solo entrepreneurs in the US. General Catalyst, QED Investors and Gradient Ventures led the Series B, with Intuit Ventures, FT Partners and SemperVirens. Collective plans banking and insurance products. Collective

California based Foundry Robotics raised a $100 million Series A led by Marlinspike, with Khosla Ventures, Red Glass Ventures, SE Ventures, Wing VC and Axiom Ventures. It will build more robotic assembly facilities, which use its FoundryOS to run production of products such as unmanned systems, satellites and batteries. Shipbuilder HII is a customer. Foundry Robotics

Solink, founded in Ottawa and based in Toronto, raised more than $100 million led by Vertu Capital, with Goldman Sachs, Golub Growth and Comcast Ventures. Its AI analyzes security camera footage for retailers, restaurants and warehouses. Solink

New York based Crux Labs raised more than $6 million in a round led by Benchstrength and Misneach, with Analog Ventures and Pareto Holdings. It tests AI agents across production traffic using its customers' expert knowledge. Crux Labs

Voice AI company Deepgram is buying Daily, the developer of the open source Pipecat framework for real time voice and video, TechCrunch reports. Terms were not disclosed. Deepgram says Pipecat will stay open source, and it introduced the open Real Time Context Standard for voice agents. Deepgram and Daily

Mintive raised a $7 million seed round led by Cowboy Ventures for managing AI agent identities and permissions, Dream came out of stealth with $17 million led by Andreessen Horowitz for AI models that generate building designs, and SpendRule raised $4 million led by Mark Cuban's Radical Investments to catch hospital supplier overcharges. Mintive · Dream · SpendRule

CoverGo, based in San Francisco and Hong Kong, raised $6 million in the first close of a Series B1 round led by Betatron Venture Group for its AI insurance platform. CoverGo

Dulo, co-founded by Sebastian Thrun and Adrian Aoun, raised a $130 million seed round for foundation models for hardware design, backed by Eric Schmidt, John Doerr, DST Global and others. Dulo

Thrive Holdings, backed by OpenAI, agreed to buy ADT in an all cash deal valued at $7.6 billion in equity and $15.3 billion in enterprise value, $10 a share, to close in the first quarter of 2027. Thrive and ADT

Coveo acquired Anglera, adding its AI product data enrichment to Coveo Commerce to improve product discovery on ecommerce sites and in AI shopping assistants; terms were not disclosed. Coveo

Legal

Harvey bought Aida, an AI sales execution company founded by Glean co-founder Piyush Prahladka and Atish Das Sarma, its fifth acquisition in 2026, Artificial Lawyer reports. Aida's founders and its search and machine learning engineers will help law firms organize internal knowledge into ontologies that give Harvey's agents and models more context. Harvey and Aida

Bloomberg Law gives legal professionals federal and state court dockets, filing information, alerts and litigation intelligence inside Claude through a Model Context Protocol connector. Bloomberg Law

LegalZoom launched a free AI assistant that gives small business owners guidance from their records, also as apps in ChatGPT, Claude, Copilot, Gemini and Perplexity. LegalZoom

Ropes & Gray and OpenAI built a tool that reviews large document sets, groups commercial contracts by issue and drafts diligence reports; lawyers check the documents and outputs before clients get them. Ropes & Gray

Data centers and chips

Telehouse and KDDI completed the purchase of the Inventory Estate at East India Dock in East London from LaSalle Investment Management and Trilogy Real Estate for more than £220 million. The site will expand Telehouse's London Docklands campus for a planned 65MW data center. Telehouse

Indian conglomerate The Chatterjee Group launched Cortex DC, an AI subsidiary that plans 2GW of data center capacity across India. The first facility should run within 18 months of the start of construction, with 51% of its energy from renewable sources, DCD reports. Cortex DC

Bull signed a memorandum of understanding to join the Czech AI Gigafactory consortium led by České Radiokomunikace, which is bidding under the European AI Gigafactories program. The deal covers hardware, software and possible server manufacturing with Foxconn and involves more than 200 Bull AI engineers. The facility could need CZK 100 billion in investment. Czech gigafactory

Daikin expects its North American data center cooling business to pass $3.3 billion by fiscal 2030. Its 2026 US Business and Innovation Report says it will raise cooling production capacity to 2.5 times 2024 levels, invest in a San Diego Data Center Hub and build a $163 million research test lab for new cooling technology. Daikin

Short seller Culper Research claims Cerebras Systems' wafer scale chips last 1.4 to 2.3 years, less than the three to five years stated in draft IPO filings. Culper also disputes Cerebras' claims about its chip materials and says dummy operations cause high power use. Cerebras was asked for comment, DCD reports. Cerebras

US intelligence agencies reportedly surveilled Megaspeed executive Alice Huang over suspected ties to China, The New York Times reports, citing anonymous sources; Nvidia denied the concerns. Alibaba Cloud opened its first Netherlands region with two Amsterdam data centres. Megaspeed · Alibaba Cloud

The Lansing Economic Area Partnership and Lighthouse Sustainability propose a 700 MW data center campus near Lansing, Michigan, during a 12 month township moratorium; Delhi Township officials say they were not consulted. A University of Wyoming survey found 70% of residents oppose more large AI data centers. Michigan · Wyoming survey

Health

Google is opening a research hub in Singapore and expanding its health AI work with local researchers, universities, public agencies and healthcare providers. The work includes foundational AI models, tests of an AI co-clinician, predictive insights for population health and preventive care tools. Google.org and Google DeepMind will give $7 million to the Philanthropy Asia Alliance's Health for Human Potential Coalition, including $1 million for the PathGen platform. Google in Singapore

A Pew Research Center survey found that US adults are open to AI handling tasks such as translating for patients and scheduling, but prefer human review. Sixty percent are uncomfortable with AI supporting diagnoses and treatments, and 46% oppose any AI role in health insurance claim decisions. Pew survey

Kipu Health made agentic AI in Kipu Helix generally available. Its Utilization Review Agent prepares clinical evidence, finds documentation gaps and drafts medical necessity reviews for behavioral health providers, and a person must approve every submission to a payer. Kipu says the system draws on more than 2 trillion words of clinical data and has prepared 23,000 continued stay reviews. Kipu Health

The 53 member states of WHO's European region adopted their first roadmap for AI in health care, covering 2026 to 2033, and backed a new WHO office on AI. WHO Europe

Also

Lone Star Liberty, a Texas super PAC that backs Republican Senate candidate Ken Paxton and received $10 million from Elon Musk, is using conversational AI calls to accuse Democratic candidate James Talarico of taking money from data center interests, CNN reports. The calls use an AI voice named Jennifer that answers voters in real time. AI robocalls

South Korea's Education Ministry plans to seek criminal investigations and trials for students who use AI powered glasses to cheat on the College Scholastic Ability Test. It also plans to extend retest bans from one year to up to three years, place specialists at all 1,312 test sites and assess radio frequency detectors for future exams. CSAT cheating

SETDA's 2026 State EdTech Trends Report found that 79% of state education leaders rank AI as a top technology priority and that 23% of states now require AI guidance under state law. SETDA report

An ActivTrak Productivity Lab analysis of 9,140 knowledge workers at 467 enterprises found that AI users worked 8.4 to 9.1 hours a day, against 7.6 hours for nonusers. Heavy AI users switched applications more often and had shorter focus sessions. ActivTrak study

The George Washington University is a founding partner in Hugging Face's Project ProtoStar, which works with research institutions to make publicly funded datasets, models and code easier for people and AI systems to find and reuse. GW's Open Source Program Office will help researchers share work through Hugging Face. Project ProtoStar

A broad majority in the Dutch lower house backs a motion to ban generative AI for primary school pupils, while still allowing lessons about AI. Dutch schools

McGraw Hill will bring its course content into Microsoft 365 Copilot through an MCP plugin, with a limited beta in early 2027. McGraw Hill

New York lawmakers introduced RAISE Act 2.0, which would require independent safety audits of AI developers and let the state limit computing use or halt dangerous deployments. RAISE Act 2.0

A UK Government study estimates AI revenues in the UK rose 41% to £33.6 billion in 2025, across 10,488 AI companies employing about 130,400 people. UK AI sector

Dartmouth provost Santiago Schnell resigned after a faculty committee found he used AI in academic and public writing without the required disclosure. Dartmouth

Representatives Robert Garcia and Jamie Raskin asked the White House to preserve President Trump's chatbot prompts and responses after reports that he consulted Grok about US military action in Venezuela. Grok records

Scale AI and Hugging Face released FineART, an open robotics dataset of 40,543 episodes and 1,718 hours across 151 tasks, with a LeRobot model that plans subtasks before acting. FineART

Canada's federal, provincial and territorial privacy regulators called for responsible AI in health care and research, with stronger privacy protections and governance. Canada

A UK government survey of 3,503 businesses found 32% used at least one AI technology in 2026, up from 16% in 2025. UK adoption

UNC-Chapel Hill stopped using an AI tool to score grammar and writing in admissions essays, because students' own AI use made the ratings less useful. UNC admissions

The Document Foundation says keeping AI out of LibreOffice's default install is about built-in features: contributors may use LLM tools if they disclose it, understand the code and pass human review. LibreOffice

A POGO investigation reports that retired generals and former defense officials helped Rhombus Power promote predictive AI tools to the US military that often failed to perform as advertised, according to former employees and military assessments; SOCOM says it uses no Rhombus tools now. Rhombus Power

Rivian is licensing parts of its AI voice assistant to Volkswagen Group through their joint venture; it could first appear in the low cost ID EVERY1. Rivian