Daily AI / Tech / VC Briefing: Wednesday 28 October 2026

Good evening, here's today's briefing. Anthropic says an unreleased Claude model found and chained real security flaws into working attacks on Chrome, Safari, Android, Windows and Linux. Its IPO filing keeps giving up new detail, from a retail share allocation to an FTC inquiry and Amazon's outsized stake. OpenAI is reportedly set to buy Google's AI chips and has pulled Oracle's software into ChatGPT, while Microsoft and Alphabet report tonight. Cognition and Cursor made product moves, and London's hyperexponential is being bought for $1.7 billion.
LEAD: Anthropic's unreleased Claude model built working exploit chains for Chrome, Safari, Android, Windows and Linux
Background: Anthropic already restricts its most capable cyber model, Mythos, to vetted customers, and rivals are packaging these models into security products (yesterday's briefing covered Cognition's Devin Security). Until now, chaining several separate bugs into one working attack has been the work of a few specialist teams. What happened: Anthropic's Frontier Red Team says an unreleased internal Claude model found flaws and chained them into five proof-of-concept exploits. In Chrome, four bugs chained together ran code outside the browser's sandbox with a single click on stock Chrome on a Mac; Anthropic says it reported 62 Chrome flaws that were fixed as of Chrome 156, all rated critical or high. In Safari, a chain switched off the rule that stops one website reading another's data. Other chains copied data off a locked Android phone over USB, crashed Windows servers over HTTP/3, and gave root access to Linux servers running a multipath TCP listener with no login. Every flaw was reported first and is patched. The Hacker News notes that how much each chain the model built on its own rests on Anthropic's word, and the model isn't named. Why it matters: It's the clearest public evidence yet that AI can do top-tier offensive security work. That raises the stakes for patching speed everywhere, strengthens the case for restricted access to cyber models, and lands two weeks before Anthropic's planned listing. Sources: Anthropic · The Hacker News
Anthropic IPO: JPMorgan to lead a big retail slice, an FTC inquiry, and Amazon as the biggest winner
Background: Anthropic made its S-1 public on Monday and is targeting a listing in the week of 16 November. What happened: Bloomberg reports JPMorgan will lead sales to US individual investors, with Anthropic weighing 15% to 20% of the offering for retail buyers; Citigroup will coordinate international retail. Investors now put fair value at about $1.4 trillion to $1.5 trillion, possibly $1.75 trillion. The filing discloses an FTC civil investigative demand sent this month covering consumer protection, privacy and product safety, roughly 94 pages of risk factors, and a pause of more than a month in pre-training its next frontier model while it tightened protection of its model weights. TechCrunch's read is that its venture investors will make billions, but Amazon gains far more: its $6.2 billion of convertible notes were valued at $97.9 billion on its own books, it put in another $10 billion of preferred shares, and Anthropic has committed to spend at least $110 billion on AWS over 10 years. Why it matters: A large retail allocation copies the SpaceX playbook and puts the AI trade directly in ordinary investors' hands. The FTC inquiry and the training pause are the sort of risks public investors haven't had to price for a frontier lab before. Sources: Bloomberg via Moneycontrol · InvestmentNews · TechCrunch
OpenAI to buy Google's AI chips, brings Oracle into ChatGPT, and faces a new lawsuit, as Microsoft and Alphabet report tonight
Background: OpenAI is one of Nvidia's biggest customers and plans to deploy its own Broadcom-designed chip by year-end. It already sells its models through Google Cloud. What happened: OpenAI plans to buy Google's TPU chips, The Information reported and Reuters followed; terms weren't disclosed and neither company commented. Separately, OpenAI and Oracle expanded their deal: Oracle is building an "AI Workspace" so its business apps can be used inside ChatGPT, an "AI Security Mesh" that uses OpenAI cyber models to find and fix vulnerabilities, a small-business bundle, and an App Composer tool built on Codex. Oracle says it already has 130,000 active ChatGPT users and more than 95,000 active Codex users internally. On the legal side, the parents of a 31-year-old woman with epilepsy have sued OpenAI, alleging GPT-4o helped her taper her seizure medication without telling her doctors before her death in February; these are allegations, not findings. Microsoft (reporting its new Agents and Infra segment for the first time) and Alphabet (Google Cloud expected up about 90% to 95%) both report after the US close tonight. Why it matters: OpenAI buying from a direct rival shows how tight compute still is, and it is another win for Google's chips against Nvidia. The Oracle deal is OpenAI pushing ChatGPT as the front door to enterprise software, the same ground Anthropic and Microsoft are fighting over. Sources: Reuters via BusinessLine · OpenAI: Oracle · TechCrunch: lawsuit
Competitor moves: Cognition's tougher coding benchmark, Cursor's Annotate, and SpaceX's subscription push
Background: Coding-agent companies increasingly compete on how they are measured, not just on features, and SpaceXAI now bundles Grok, Grok Bot, Cursor and X Premium into one plan called The X Subscription. What happened: Cognition launched FrontierCode 2.0, a benchmark built from real open-source repositories by their own maintainers, graded on whether a maintainer would actually merge the agent's change. Cognition says the best model scores just 45%, with GPT-6.1 Sol the most cost-efficient at 39% for $1.03 a task, and Artificial Analysis has already folded it into its main model index. Cursor added Annotate to its design mode, so developers can click, draw or speak on their running app and those marks go to the agent as instructions. SpaceXAI launched a referral programme for The X Subscription, whose five tiers run from free to $200 a month with Grok usage shared across Grok and Cursor. Why it matters: By defining the "would you merge this?" test, Cognition is trying to set the bar its rivals get judged by. Cursor is going after front-end work where pointing beats describing, and SpaceX is using referrals to pull Cursor users into its wider bundle. Sources: Agentdex roundup · Cursor on X · PYMNTS
Deal: Guidewire buys London's hyperexponential for $1.7 billion to push AI underwriting
Background: hyperexponential, or hx, is a London start-up whose software helps commercial insurers price risk. Its customers include Aviva, Allianz and Beazley. What happened: Insurance software maker Guidewire agreed to buy hx for $1.7 billion in cash plus up to $400 million more if revenue targets are hit, funded by a $1 billion loan and cash. hx says its platform already supports underwriting decisions on more than $75 billion of annual premium. Guidewire pitches the deal as "agentic underwriting", with AI agents assessing and pricing risk alongside human underwriters. Why it matters: It's a big cash exit for a UK AI company and a sign that incumbents are paying up for industry-specific AI agents rather than building them. Sources: Guidewire
VC: Disruptive closes a $10 billion fund for about 10 big bets, plus agent and fraud rounds
Background: Disruptive is a Texas growth investor that has never before raised a formal flagship fund. What happened: Disruptive announced its first formal fund at $10 billion, the largest venture fund in Texas history, planning cheques of up to $1 billion each into roughly 10 late-stage companies in AI, AI infrastructure, defence and enterprise software. Today's rounds include Fragment, which launched with $56 million (Series A led by 8VC, seed co-led by General Catalyst) for procurement agents that charge per job done rather than per seat; Casap, a $56 million Series B led by Base10 for AI that helps banks fight card-dispute fraud; and London's Circuit Mind, $25 million in total including an $18.5 million Series A led by Norwest, for AI that helps engineers design electronic circuits. Why it matters: Mega-funds concentrating on a handful of AI winners are the late-stage mirror of the Anthropic IPO, while early money keeps flowing to agents that are paid for completed work. Sources: TechCrunch: Disruptive · Axios · Fragment · Fortune: Casap · Circuit Mind
No major new hires at coding or AI-agent companies surfaced today beyond those already covered this week.