October 24, 2025

This Week in AI (Oct 18–24, 2025): The Browser Wars Heat Up, Enterprise AI Goes Mainstream, and Compute Becomes Strategy

The race to control how people browse the web intensified dramatically this week as three major players released AI-native browsers designed to fundamentally change how we interact with the internet.

By Jay Cadmus

Browser windows glowing on a dark desk at night.

Agentic Browsers: The New Frontier

The race to control how people browse the web intensified dramatically this week as three major players released AI-native browsers designed to fundamentally change how we interact with the internet.

ChatGPT Atlas: OpenAI's Chrome Challenger

On October 21, OpenAI unveiled ChatGPT Atlas, a Chromium-based browser with ChatGPT built directly into the browsing experience. Available now for macOS (with Windows, iOS, and Android versions coming soon), Atlas represents OpenAI's bid to move beyond being just a chatbot and become the primary interface for how users experience the web.

Atlas features a ChatGPT sidebar that allows users to ask questions about any page they're viewing, summarize content, compare products, and analyze data from websites. The browser includes an optional agent mode that can perform tasks autonomously; researching topics, filling out forms, and even shopping on behalf of users (currently limited to Plus and Pro subscribers).

One of the most significant innovations is browser memories, an optional feature that allows ChatGPT to remember context from sites visited (past job listings, hotel searches, or research topics) to provide more personalized future responses. OpenAI emphasizes user control: memories can be reviewed, archived, or deleted at any time, and an incognito mode is available for AI-free browsing.

CEO Sam Altman framed the launch as rethinking what a browser can be: "We think AI represents a rare once-a-decade opportunity to rethink what a browser can be about and how to use one and how to most productively and pleasantly use the web. Tabs are great but we haven't seen a whole lot of innovation since then."

The browser is a direct challenge to Google Chrome's 72% market share and puts OpenAI in competition with Google's core business: controlling the gateway to the internet. With over 800 million weekly ChatGPT users, Atlas has immediate distribution potential that could reshape the browser landscape.

Perplexity Comet Goes Free

Perplexity AI announced on October 2 that its Comet browser is now available worldwide for free, after initially launching in July exclusively to Perplexity Max subscribers at $200 per month. The waitlist reportedly reached "millions" of users before the broader rollout.

Comet is designed to function as a personal assistant capable of web searching, managing tabs, composing emails, online shopping, and more. The browser processes AI locally rather than in the cloud, offering privacy benefits and ensuring users' site visits aren't used to train AI models. Users can highlight any text, sentence, or image and Comet will discuss it, summarize articles, write RSVPs, or organize travel itineraries.

Perplexity is positioning Comet as superior to rival offerings from Google, OpenAI, and Anthropic. The company has already formed revenue-sharing partnerships with major publishers including CNN, Condé Nast, The Washington Post, Los Angeles Times, Fortune, Le Monde, and Le Figaro to address earlier content plagiarism concerns.

Looking ahead, Perplexity indicated more features are in development, including a mobile version of Comet and a new capability called Background Assistant that can handle multiple tasks simultaneously and asynchronously.

BrowserOS: The Open-Source Alternative

Y Combinator-backed BrowserOS emerged this week as an open-source, AI-native browser focused on privacy and local execution. With 4,300+ GitHub stars, 25,000+ downloads, and an active Discord community of 1,000 members, BrowserOS represents a grassroots alternative to corporate AI browsers.

The key differentiator: BrowserOS runs locally with users' logged-in sessions, solving a critical limitation of cloud-based solutions like ChatGPT Operator or other remote agent systems that can't access Gmail, LinkedIn, or company tools requiring authentication. Users describe tasks in plain English — like "find software engineers from my LinkedIn requests and add them to a Google sheet" — and the AI clicks, types, and navigates on their behalf.

BrowserOS positions itself as a privacy-first alternative to Comet, Dia, and Chrome, particularly appealing to enterprises concerned about data security and compliance. The project aims to deliver the browser's "Cursor moment", the transformative productivity boost that AI coding tools brought to developers, but extended to the billion knowledge workers who spend half their time on repetitive browser tasks.

The Security Challenge

As agentic browsers proliferate, cybersecurity experts are raising urgent warnings. Researchers at Brave Browser demonstrated how AI-powered browsers can fall prey to prompt injection attacks, where malicious webpages embed hidden instructions that the AI unwittingly follows.

In one demonstration, a hacker placed invisible prompts in a Reddit comment. When Perplexity Comet was asked to summarize the page, it autonomously accessed the user's Perplexity login, grabbed the email address, read the verification email from Gmail, and posted the login credentials as a Reddit comment, all without the user realizing. The entire account takeover happened simply because someone visited a webpage.

The "scariest aspect," according to Brave researchers, is that an agentic browser hijacked by a malicious site could access banking, work email, or other sensitive accounts using the user's authenticated privileges. Because AI agents operate with the same permissions as users, a successful exploit can be catastrophic.

OpenAI, Microsoft, and Perplexity will need robust guardrails — including sandboxing, permission controls, and constant security patching — to prevent AI assistants from becoming attack vectors. Expect new browser settings for AI consent flows, corporate policies limiting agent permissions, and an arms race between AI security teams and malicious actors.

Enterprise AI Infrastructure

This week marked a turning point as AI moved to mission-critical enterprise infrastructure.

ChatGPT Company Knowledge

OpenAI introduced Company Knowledge for Business, Enterprise, and Education plans on October 23, allowing ChatGPT to connect directly to internal knowledge sources with proper permissions and citations. Organizations can set up connectors to Slack, SharePoint, Google Drive, GitHub, email, and other apps.

Once enabled, employees can ask questions that span multiple systems. Have it "Summarize the latest customer feedback from our Slack channel and Google Form responses", and ChatGPT retrieves relevant snippets, synthesizes them into a single answer, and provides inline citations with source links for verification and audit trails.

Critically, Company Knowledge respects existing access controls: ChatGPT can only retrieve data a user already has permission to see. It does not train on company data by default, avoiding the risk of inadvertently leaking sensitive information into the model. Enterprise admins must approve each integration and can set role-based access policies with visibility into usage logs via an audit API.

The feature launched with a brief performance incident that was transparently reported and resolved, demonstrating OpenAI's commitment to reliability for enterprise deployments.

Microsoft Copilot Fall 2025 Release

Microsoft held a major event on October 23 to unveil 12 new Copilot features, positioning the platform as practical AI infrastructure rather than experimental technology. The updates span collaboration, personalization, integration, and specialized use cases:

Groups — Collaborative Copilot sessions allowing up to 32 participants to brainstorm, co-author, or plan together via a shared link, with Copilot summarizing discussions and tracking actions.

Memory & Personalization — Long-term memory allowing Copilot to remember key information like training plans, important dates, and goals at the user's request, with full user control.

Connectors — Integration with Google Drive, Dropbox, Gmail, and Google Calendar for natural-language searches across accounts.

Copilot Mode in Edge — Transforms Microsoft Edge into an "AI browser" that summarizes, compares, and executes web actions via voice commands.

Mico — A newly introduced avatar character that provides expressive feedback and emotional responses, serving as a cohesive user experience layer.

Real Talk — A conversational mode that provides constructive pushback, fostering a Socratic approach to problem-solving.

Copilot for Health — Health-related information sourced from reputable medical entities like Harvard Health.

The updates demonstrate Microsoft's strategy to make Copilot ubiquitous across Windows, Edge, Office, and enterprise data sources. Notably, Microsoft is shifting from reliance on partner OpenAI to using its own in-house models like MAI-V-1, MAI1-Preview, and MAI-Vision-1, signaling a competitive divergence from its longtime AI partner.

Google Gemini Enterprise & Computer Use

Google launched Gemini Enterprise, an agentic platform with AI Agent Finder tools to help businesses discover and procure AI agents. The company also released the Gemini 2.5 Computer Use model via API, which outperforms leading alternatives at browser and mobile tasks with lower latency.

The Computer Use model powers agents capable of interacting with user interfaces by clicking, typing, and scrolling. This enables tasks like filling forms, manipulating dropdowns, and operating behind logins. This capability is crucial for building general-purpose agents that can handle tasks requiring direct GUI interaction rather than just API calls.

Google also unveiled major Veo 3.1 video generation updates, automated Canvas presentation building, and voice assistant integration for Google TV as part of its October Gemini Drop.

Compute at Scale: The Infrastructure Arms Race

Anthropic's 1 Million TPU Deal with Google

On October 23, Anthropic announced a massive expansion of its use of Google Cloud TPUs, with plans to access up to 1 million TPU chips worth tens of billions of dollars and over 1 gigawatt of capacity expected to come online in 2026.

The deal allows Anthropic to train next-generation Claude models using Google's Ironwood-era TPU hardware, which the company cited for superior price-performance and efficiency compared to alternatives. Anthropic reiterated a multi-platform compute strategy, continuing to use Amazon Trainium and NVIDIA GPUs alongside Google TPUs.

Securing dedicated AI compute at this scale is now table stakes for frontier model roadmaps. The agreement diversifies beyond NVIDIA's GPU monopoly and tightens Google's position as a credible alternative for large-scale AI training. It also highlights the critical role of procurement, energy, and data center siting strategies in shaping competitive timelines. Companies that lock in compute capacity years in advance have a structural advantage in the AI race.

The deal's scope, 1 million chips and multiple gigawatts of power, underscores the insatiable demand for AI hardware and the capital-intensive nature of frontier AI development.

Meta's Record $27 Billion Data Center Deal

In Wall Street's biggest private-credit deal ever, Meta Platforms and Blue Owl Capital announced a $27 billion joint venture for Meta's Hyperion data center campus in Louisiana. The deal signals a pivot in how Big Tech is funding its AI ambitions, moving toward alternative financing structures rather than traditional capital markets.

The Hyperion campus will support Meta's AI infrastructure needs as the company integrates AI across Facebook, Instagram, WhatsApp, and its metaverse initiatives. The financing structure reflects the massive capital requirements of AI-scale data centers and the creative financial engineering required to fund them.

Platform Policy as Competitive Weapon

WhatsApp Bans Third-Party AI Chatbots

Meta updated WhatsApp's Business API policy on October 18 to ban general-purpose AI chatbots from its platform, effective January 15, 2026. The move affects AI companies including OpenAI, Perplexity, Khosla Ventures-backed Luzia, and General Catalyst-backed Poke, which currently use WhatsApp to deliver conversational services to millions of users.

The updated policy states that "AI Providers" are "strictly prohibited from accessing or using the WhatsApp Business Solution" for distributing general-purpose AI assistants. Meta clarified that business-specific customer service bots — like travel booking or order processing — can continue operating.

Meta's rationale is that WhatsApp Business API was designed for businesses serving customers, not as a platform for chatbot distribution. The company said chatbot use cases placed unexpected burden on its systems with increased message volume and required different support than it was prepared to provide.

OpenAI confirmed ChatGPT will wind down on WhatsApp and published instructions for users to link accounts and continue chats in ChatGPT apps. With 50+ million users affected, the transition represents a significant disruption.

The move underscores how platform policy is becoming a competitive lever. For product teams relying on third-party messaging surfaces, this creates channel risk and a push toward owned apps/browsers or enterprise integrations. For users, AI assistants will increasingly be native to platform ecosystems, with Meta AI on WhatsApp and ChatGPT in its own apps.

X Algorithm Shifts to Grok-Powered AI

X (formerly Twitter) owner Elon Musk announced on October 18 that X's feed algorithm will soon be fully powered by xAI's Grok system. Musk stated: "The X recommendation system is evolving very rapidly. We are aiming for deletion of all heuristics within 4 to 6 weeks. Grok will literally read every post and watch every video (100M+ per day) to match users with content they're most likely to find interesting."

The update should help address X's new user and small account problem, where quality posts from unknown accounts receive little visibility. By leveraging Grok's language understanding, X aims to surface better content regardless of account size or follower count.

Interestingly, Musk's comment that X sees 100 million original posts daily (the same figure reported in 2023) suggests X's posting volume hasn't grown in two years; a notable data point on platform stagnation despite AI-driven changes.

Vertical AI: Life Sciences Acceleration

Anthropic announced significant advancements in tailoring Claude for life sciences research on October 19, rolling out scientific connectors, Agent Skills, and improved model performance for drug discovery and clinical work.

Claude Sonnet 4.5 scores 0.83 on Protocol QA (a benchmark testing understanding of laboratory protocols), exceeding the human baseline of 0.79. The model now integrates with scientific platforms including:

10x Genomics — Single-cell and spatial analysis capabilities through plain English conversation, lowering barriers for non-computational researchers.

Benchling — Lab notebook and genomic database integration for streamlined research workflows.

Komodo Health — Healthcare analytics designed for regulated environments with transparent, auditable solutions.

Early pharma partners report dramatic productivity gains. Novo Nordisk reduced a clinical documentation process from 10 weeks to 10 minutes using Claude. The focus is on assistive roles rather than autonomous science, with safety checks including reduced hallucinations, audit trails for regulatory compliance, and grounded sourcing.

Anthropic's partnership approach, working with established scientific software providers rather than building direct-to-researcher tools, reflects a pragmatic strategy for penetrating highly regulated industries that demand provenance and auditability.

Notable Developments

OpenAI acquires Sky: OpenAI acquired Software Applications Inc., maker of Sky, a Mac-native AI interface that "sees" screen content and can take actions in apps. The team joins ChatGPT, foreshadowing tighter desktop integration, especially on macOS.

PBS metadata automation: Media giant PBS used Amazon Bedrock AI to auto-tag 700,000+ media assets in months, dramatically improving content discoverability across digital platforms.

Mondelez AI-generated ads: Oreo-maker Mondelez is investing in AI-generated advertising to cut marketing content costs by up to 50%, with plans to scale from test-and-learn phases to broader deployment.

Anthropic Seoul expansion: Anthropic announced plans to open a Seoul office in early 2026, citing 10x APAC revenue run-rate growth year-over-year.

Salesforce-Anthropic partnership expansion: Salesforce and Anthropic expanded their strategic partnership to deliver Claude models in Agentforce for regulated industries including financial services, healthcare, cybersecurity, and life sciences. CrowdStrike and RBC Wealth Management are early adopters.

Resistant AI Series B: Czech Republic-based fraud prevention fintech Resistant AI raised $25 million in Series B funding on October 12 to expand anti-fraud technology capabilities.

Takeaways & Near-Term Actions

Agentic interfaces are moving into browsers and desktops. Expect rapid experimentation with task automation and page-aware assistance, accompanied by new controls for privacy, memory, and authorization. Organizations should pilot agentic workflows in controlled domains before broad deployment.

Enterprise chat is shifting from "paste-in context" to governed, connector-driven knowledge. The shift raises the bar for adoption in regulated industries. IT teams should reassess data governance, connector permissions, and audit-trail capabilities before enabling company-wide access.

Compute procurement is strategy. Locking in at-scale capacity signals model cycle cadence and vendor diversification. Companies without secured compute pipelines for 2026–2027 may face competitive disadvantages as frontier labs race ahead.

Platform distribution risk is real. WhatsApp's policy change underscores the strategic value of owning primary user interfaces. Product teams relying on third-party platforms should develop contingency plans and owned distribution channels.

Security for agentic systems is unsolved. Prompt injection, credential leakage, and autonomous action risks require new frameworks. Organizations deploying AI agents should implement strict permission boundaries, human-in-the-loop verification for sensitive actions, and continuous monitoring.

Vertical AI is accelerating in regulated industries. Life sciences, healthcare, and financial services are seeing purpose-built AI solutions with compliance, auditability, and safety built in from the start. Buyers should prioritize vendors with domain expertise and regulatory track records.

References

Anthropic. (2025–10–19). Claude for Life Sciences.

https://www.anthropic.com/news/claude-for-life-sciences

Anthropic. (2025–10–22). Expanding our use of Google Cloud TPUs and Services.

https://www.anthropic.com/news/expanding-our-use-of-google-cloud-tpus-and-services

CNN. (2025–10–22). The battle for the future of the internet is underway.

https://www.cnn.com/2025/10/22/tech/openai-chatgpt-atlas-browser-google-chrome-ai

CNBC. (2025–10–02). Perplexity AI rolls out Comet browser for free worldwide.

https://www.cnbc.com/2025/10/02/perplexity-ai-comet-browser-free-.html

Fortune. (2025–10–23). Cybersecurity experts warn OpenAI's ChatGPT Atlas is vulnerable to attacks that could turn it against a user.

https://fortune.com/2025/10/23/cybersecurity-vulnerabilities-openai-chatgpt-atlas-ai-browser-leak-user-data-malware-prompt-injection/

Futurism. (2025–10–07). Why AI Browsers Like Perplexity's Comet Could Make the Web Riskier.

https://futurism.com/artificial-intelligence/researchers-severe-vulnerabilities-ai-browser-comet

Google DeepMind. (2025–10–06). Introducing the Gemini 2.5 Computer Use model.

https://blog.google/technology/google-deepmind/gemini-computer-use-model/

Microsoft. (2025–10–22). Human-centered AI.

https://www.microsoft.com/en-us/microsoft-copilot/blog/2025/10/23/human-centered-ai/

Morningstar. (2025–10–23). Meta, Blue Owl and AI: Here are the details of Wall Street's biggest private-credit deal ever.

https://www.morningstar.com/news/marketwatch/20251023242/meta-blue-owl-and-ai-here-are-the-details-of-wall-streets-biggest-private-credit-deal-ever

NBC News. (2025–07–15). AI-powered web browsers want to help people save time. But are they effective?

https://www.nbcnews.com/tech/tech-news/ai-powered-web-browsers-want-help-people-time-are-effective-rcna218039

OpenAI. (2025–10–20). Introducing ChatGPT Atlas.

https://openai.com/index/introducing-chatgpt-atlas/

Perplexity. (2025–10–21). Introducing Comet: Browse at the speed of thought.

https://www.perplexity.ai/hub/blog/introducing-comet

Reuters. (2025–10–21). OpenAI launches AI browser Atlas in latest challenge to Google.

https://www.reuters.com/technology/openai-unveils-ai-browser-atlas-2025-10-21/

Reuters. (2025–10–23). Anthropic to use Google's AI chips worth tens of billions to train Claude chatbot.

https://www.reuters.com/technology/anthropic-expand-use-google-clouds-tpu-chips-2025-10-23/

Reuters. (2025–10–23). Microsoft introduces new Copilot features such as collaboration, Google integration.

https://www.reuters.com/technology/microsoft-introduces-new-copilot-features-such-collaboration-google-integration-2025-10-23/

Reuters. (2025–10–24). Oreo-maker Mondelez to use new generative AI tool to slash marketing costs.

https://www.reuters.com/business/media-telecom/oreo-maker-mondelez-use-new-generative-ai-tool-slash-marketing-costs-2025-10-24/

Salesforce. (2025–10–13). Salesforce, Anthropic Expand Partnership Supporting Regulated Industries.

https://www.salesforce.com/news/press-releases/2025/10/14/anthropic-regulated-industries-partnership-expansion-announcement/

TechCrunch. (2025–10–17). WhatsApp changes its terms to bar general-purpose chatbots from its platform.

https://techcrunch.com/2025/10/18/whatssapp-changes-its-terms-to-bar-general-purpose-chatbots-from-its-platform/

TechCrunch. (2025–10–19). Anthropic brings Claude Code to the web.

https://techcrunch.com/2025/10/20/anthropic-brings-claude-code-to-the-web/

TechCrunch. (2025–10–22). OpenAI's Atlas is more about ChatGPT than the web.

https://techcrunch.com/2025/10/22/openais-atlas-is-more-about-chatgpt-than-the-web/

TechCrunch. (2025–10–23). OpenAI buys Sky, an AI interface for Mac.

https://techcrunch.com/2025/10/23/openai-buys-sky-an-ai-interface-for-mac/

TechRadar. (2025–10–22). Meta is booting ChatGPT and other non-Meta AI assistants from WhatsApp.

https://www.techradar.com/ai-platforms-assistants/meta-will-ban-rival-ai-chatbots-from-whatsapp

VentureBeat. (2025–10–23). Microsoft Copilot gets 12 big updates for fall, including new AI assistant character Mico.

https://venturebeat.com/ai/microsoft-copilot-gets-12-big-updates-for-fall-including-new-ai-assistant

VentureBeat. (2025–10–23). OpenAI launches company knowledge in ChatGPT letting you access your firms.

https://venturebeat.com/ai/openai-launches-company-knowledge-in-chatgpt-letting-you-access-your-firms

Washington Post. (2025–10–22). ChatGPT just came out with its own web browser. Use it with caution.

https://www.washingtonpost.com/technology/2025/10/22/chatgpt-atlas-browser/

Y Combinator. (2025–09–18). BrowserOS: The Open Source Agentic Browser.

https://www.ycombinator.com/companies/browseros