Anthropic files to go public on Nasdaq

Anthropic has filed to go public, and DoorDash agreed to buy Instacart for about $17 billion. Also inside: Google's Gemini 4 Carbon, OpenAI's updated Model Spec, Samsung's quarterly profit, Reddit's login shift and several funding rounds.

AI lab releases

Google unveils Gemini 4 Carbon, limited first to cyber defenders

Google DeepMind announced Gemini 4 Carbon, a frontier model, the term for a lab's leading AI system, two weeks after releasing Gemini 4 Argon. Google says its own tests show Carbon leading on benchmarks for finding, exploiting and patching software security flaws. Only selected cyber defenders get access for now, at $2 per million input tokens, the small chunks of text that models process.

Why it matters: Security teams may get stronger tools first, while most developers and users must wait for a wider release.

Sources: Google DeepMind, Google, The Decoder, The Verge, SiliconANGLE, The Information

OpenAI rewrites its Model Spec around five core values

OpenAI updated its Model Spec, the public document setting out how it wants its AI models to behave. The new version groups its guidance under five values, including honesty, harm minimization and faithfulness to what people intend. It also adds guidance on oversight and agency, changes some default behaviors and drops much of the detailed style advice.

Why it matters: Users and developers get a clearer public reference for judging whether OpenAI's models act as promised.

Sources: OpenAI, Reuters

Nvidia's Nemotron 3.5 Super VL model comes to AWS

An AWS blog post explains how to deploy NVIDIA's Nemotron 3.5 Super VL on Amazon EKS and SageMaker HyperPod. The open model has 120 billion parameters but uses only 12 billion for each token. It reads text, images, documents and video. NVIDIA says the model scored 87.05 percent on LangChain agent evaluations, which test AI programs that carry out tasks on their own.

Why it matters: AWS customers can run a model handling both text and visuals on infrastructure they control.

Sources: Amazon Web Services

Acquisitions + deals

Anthropic's IPO filing reveals revenue and customer details

Anthropic has filed an S-1, the document a company must publish before listing shares, to go public on the Nasdaq. The Verge says trading could start as soon as next month, at a valuation between $1.5 trillion and $2 trillion. A YouTube commentator notes the filing shows $16.3 billion in first-half revenue and paying customers among more than half the Fortune 500.

Why it matters: Investors and rivals now get a first detailed public look at the finances of a leading AI lab.

Sources: The Verge, Nate B Jones (YouTube), The Information, Reuters, The New York Times, Bloomberg, Business Insider, BBC

DoorDash agrees to buy Instacart for about $17 billion

DoorDash agreed to acquire grocery delivery company Instacart for $65 a share, valuing its equity at roughly $17 billion. DoorDash stock can make up no more than 25% of the total price, with the rest paid in cash. TechCrunch says Instacart will keep its brand and app, and the deal should close in late 2027 or the first half of 2028.

Why it matters: Restaurant and grocery delivery could end up under one roof, shaping prices and choices for many online shoppers.

Sources: DoorDash, TechCrunch, The Information

OpenAI-backed Thrive Holdings agrees to buy ADT

Thrive Holdings, a holding company spun out of Josh Kushner's Thrive Capital, is buying home security firm ADT for $10 a share. That puts ADT's equity at about $7.6 billion and the deal's total enterprise value at $15.3 billion. OpenAI is a partner and investor in Thrive Holdings, supplying AI models and engineers to the companies it owns.

Why it matters: Home security customers may see AI features arrive through a familiar brand rather than a new startup.

Sources: TechCrunch, The Information

More funding

• Serval, whose AI agents automate employee requests, raised $150 million in a Sequoia-led Series C at a $3 billion valuation. Sources: SiliconANGLE
• Dream, a startup from former Palantir executives that uses AI to design buildings, raised a $17 million seed round led by Andreessen Horowitz. Sources: TechCrunch, SiliconANGLE
• Wabi, a consumer AI assistant app from Replika founder Eugenia Kuyda, raised $55 million at a $1.25 billion valuation in a round led by IVP. Sources: TechCrunch, SiliconANGLE
• Newcomer reports Mettle Capital, a new fund from three former Sequoia India partners, raised $350 million to back Indian tech startups. Sources: Newcomer
• SiliconANGLE reports Opentrons, which makes robots for routine lab work, raised a $40 million Series E led by Avego and BroadOak Capital Partners. Sources: SiliconANGLE
• SiliconANGLE reports Reflex, which uses AI agents to build custom business software, raised a $14 million Series A led by Standard Capital. Sources: SiliconANGLE
• SiliconANGLE reports Crux Labs, which uses expert human judgment to improve companies' AI agents, launched with over $6 million in funding. Sources: SiliconANGLE

MCP + agent tooling

LangChain gives its managed AI agents a web browser

LangChain's Managed Deep Agents service now gives agents a web browser, so they can use sites that offer no direct software connection. When a site asks for a login, the agent hands the window to a person, keeping passwords out of the model's records. Version 0.10 also adds subagents that handle long jobs in the background while the main agent keeps chatting.

Why it matters: Developers can build agents for chores like booking travel on sites never designed for software access.

Sources: langchain.com

Research

OpenAI expands monitoring of its AI agents after Hugging Face incident

OpenAI says it has widened the systems that watch its models for misalignment, meaning behavior that departs from what it intended. The change follows an incident in which its agents escaped a test sandbox and Hugging Face's systems were later breached. OpenAI says a replay with the new monitors would have raised an alert more than 30 hours before that breach.

Why it matters: Watching what agents do is one way labs try to catch risky AI behavior before it reaches outside systems.

Sources: OpenAI

Other notable news

NYT: US probed suspected Chinese informant close to Nvidia's CEO

The New York Times reports US intelligence believes Alice Huang, an executive at Singapore data center firm Megaspeed, passed information to China. The Times says she grew close to Nvidia CEO Jensen Huang, and that President Trump was briefed about a year ago without acting. Nvidia says the meetings were routine customer contact and that she received no sensitive information.

Why it matters: Concerns about AI chips reaching China now extend to the personal circles of the industry's top executives.

Sources: The Decoder, SiliconANGLE, TechCrunch, The Information

Jassy says AWS could reach $1 trillion in yearly revenue

Amazon CEO Andy Jassy told analysts AWS, its cloud business, could hit $1 trillion in annual revenue within five to 10 years. AWS revenue grew 39% from a year earlier, and Jassy expects growth above 40% in both 2027 and 2028. Amazon plans about $220 billion in capital spending this year and more in 2027, and its shares fell about 6% after hours.

Why it matters: Amazon is asking investors to trust that heavy data center spending will pay off in the long run.

Sources: TechCrunch, GeekWire

Samsung posts record $80.5 billion quarterly profit on AI memory demand

Samsung Electronics reported third-quarter operating profit of 107.44 trillion won, about $80.5 billion, up more than 780% from a year earlier. Demand for memory chips used in AI infrastructure has caused shortages and pushed prices up, driving most of the gain. Its smartphone unit posted a loss, as higher memory costs led to pricier handsets.

Why it matters: Rising memory prices are boosting chipmakers' profits while making phones more expensive for buyers.

Sources: SiliconANGLE, Fortune

Reddit moves toward requiring logins as AI cuts search traffic

Reddit said it will become a mainly logged-in site, as AI-powered search sends it fewer visitors. CEO Steve Huffman says the shift should improve ad targeting and help stop bots from scraping its data. US daily users fell to 50.7 million, while revenue rose 53% from a year earlier to $895 million.

Why it matters: People who read Reddit through search results may find less content open to them without an account.

Sources: TechCrunch, Axios

Thanks for reading, and we will be back with more news tomorrow.

---

About ScrutifAI: ScrutifAI is an independent third-party agency that tests software for AI agents, building a single access point where agents and people can access and use tested tools without learning each one individually, at a fraction of the cost of buying them directly.