Anthropic IPO and Cybersecurity Expansion

Welcome to Future Bytes News, your weekly briefing on AI and business. This week the AI vendor map that most of your roadmaps depend on shifted in three days. And the shift happened in two places at once, with both companies in motion at the same time. On Tuesday, June 2nd, Microsoft opened its Build Developer Conference in San Francisco. And for the first time since Microsoft started shipping AI features inside its products, the keynote was not about distributing somebody else's model. It was about Microsoft's own. Microsoft announced a coding model called Mayi Code1Flash, and it started rolling it out the same day inside GitHub Copilot, across every tier, including the free one. Microsoft says it beats Anthropics Claude Haiku model on the four coding bench marks they tested, with the biggest lead being 16 points on a benchmark called SWE Bench Pro. That benchmark measures whether a model can resolve real software engineering tasks. So this is not just a model that exists in a research paper. It is the model that starting this week is doing the auto-complete and the agentic work inside the GitHub copilot that tens of millions of developers already use. And then in the same keynote, Microsoft announced a second in-house model, its first reasoning model, called MAI Thinking One. Now this one is worth slowing down on, because the framing matters. Microsoft trained this model from scratch on commercially licensed data. They explicitly said it was not distilled from any third-party model, meaning no shortcut through OpenAI's outputs. Microsoft hired an independent rating partner called Surge to do blind side-by-side evaluations, and in those blind ratings, MAI Thinking 1 was preferred over Anthropics Claude Sonnet 4.6. On a separate coding benchmark, Microsoft says it matches Anthropix Claude Opus 4.6, which is the Frontier. So pull back for a moment. Microsoft just shipped a coding model that beats Claude Haiku , and announced a reasoning model that in blind ratings was preferred over Claude Sunnet and matches Claude Opus. The two leading model families that have been the alternative to open AI inside Microsoft's own platforms. Microsoft is now competing with both of them head to head, with models it built itself. Now here's why this matters for your business. For the past two years, the mental model most business leaders have carried about Microsoft and AI is the same one. Microsoft is the platform. OpenAI is the model. Microsoft sells distribution. Open AI sells frontier. That mental model is, as of Tuesday, out of date. Microsoft is now a frontier lab in its own right, selling its own models against Anthropic and OpenAI, inside the same products it has historically used to distribute OpenAI. That is not a small change to a vendor diligence document. That changes which box Microsoft sits in when your procurement team draws the map. The practical move on this one is straightforward. Get your head of engineering and your AI procurement lead together this month and ask one question. Who, inside your organization , owns the decision of which model your developers are actually calling when they hit tab in their editor? Because as of this week, the default fast tier inside GitHub Copilot can quietly switch from one lab's model to another, and your developers will not see any difference in the product they use. There is a second thread inside the build keynote worth flagging. Microsoft also released something they call Microsoft IQ, a context layer that grounds AI agents in your organization's data, across GitHub Copilot, Foundry, and Copilot Studio. And Microsoft started shipping the underpinnings of a Windows agent runtime. If your roadmap involves deploying AI agents against your Microsoft 365 data, the platform you will be asked to build on just became more vertically integrated. Top to bottom, Microsoft now controls the model. The context layer, the agent runtime, and the desktop. Now let me widen the frame. Same days, different company, opposite direction on a different question. The company is anthropic. And anthropic had a remarkable forty eight hours of it On Monday, June 1st, the day before Build's keynote, Enthrop ic confirmed it had confidentially filed a draft S1 registration statement with the U.S. Securities and Exchange Commission for an initial public offering. That is the legal process for going public. Anthropic has not set the share count, the price, the exchange, or the timing, but the filing puts them on the path. And it comes exactly one week after anthropic closed the largest private funding round in AI history, sixty five billion dollars at a nine hundred and sixty five billion dollar valuation. So Anthropic is now formally on the road to becoming a public company. Then on Tuesday, the same day as Microsoft Build, Anthropic announced it was expanding Project Glasswing. Now if you listen to last week's episode, you will remember Glasswing. That is Anthropic's restricted access program for Mythos, the cybersecurity AI that was deemed too dangerous to release publicly in April. Seven weeks ago, Anthropic told the world Mythos was coming to all customers. This week they expanded the restricted access program from about 50 organizations to about 200, adding roughly 150 more across more than 15 countries . And here is what changed about the recipient list. The expansion brings in critical infrastructure sectors that were not well represented in the original cohort. Power, water, healthcare, communications, hardware The named recipients include NATO, they include the European Union Cybersecurity Agency ENISA , they include Samsung, Okta, SK Heinrich, SK Telecom, and the number that should make every CISO listening sit up is this. In the eight weeks since Glasswing launched in April, Mythos has surfaced more than 10,000 high severity or critical severity software vulnerabilities, ten thousand, in eight weeks, from about fifty organiz ations. That is the volume that this capability produces when only fifty organizations have access. Pull both of these together. On Tuesday at Microsoft Build, Microsoft became a frontier model lab in its own right. The same day, Anthropic, the lab Microsoft now competes with, expanded its cybersecurity AI to NATO and the EU cybersecurity agency. The day before, Anthropic filed to become a public company. The lab you depend on for one thing just became a public company candidate. The platform you depend on for another thing just became a competing lab. A few more developments worth noting briefly. OpenAI rolled out a major upgrade to ChatGPT's memory this week, a feature called Dreaming, which lets the model automatic ally update what it remembers about you over time. OpenAI also made its frontier models and codecs generally available on Amazon Web Services, which matters because AWS has historically been Anthropic's primary cloud partner . So if your shop is on AWS and was held back from OpenAI because of the cloud, that constraint is gone. Meta started beta testing its first paid AI subscription tiers. The first sign of a major social platform trying to fund its AI build by charging consumers directly. And in Europe, the European Commission finalized its code of practice on labeling AI generated content. The rules behind Article 50 of the EU AI Act. Those rules become legally applicable on August second. If your business uses generative AI for any customer facing content in the EU, that deadline is your marketing team's deadline as much as your legal teams . For your AI roadmap, four practical questions to bring to your next planning session. First, if Mayi Code 1Flash is the default fast tier model inside GitHub CopilotNow, who on your engineering side owns the decision of which model your developers are actually calling? And when did they last review it? The lab making your code completion decisions just changed without your developers having to switch tools. Second, does your AI vendor diligence still treat Microsoft as a distribution partner for OpenAI and Anthropic? Or as a frontier lab in its own right? Because as of Tuesday, Microsoft is selling its own models against theirs . Get the diligence framework updated before the next renewal cycle. Third, if Mythos class cybersecurity capability has already surfaced 10,000 critical severity vulnerabilities in eight weeks of restricted access. What does your patch and disclosure workflow look like the day the same capability is broadly available? For defenders and for adversaries on the same day? Fourth, anthropic just filed to go public. In six to twelve months, one of your two leading AI vendors will have public financials and quarterly disclosures. And the other one is the cloud platform you may already be standardizing on. How does your vendor diligence change when the inputs you have on each vendor become asymmetric in kind ? The honest read for business leaders is direct. Your assumption that the AI labs you depend on are clearly separated from the platforms you run on has just been falsified in writing. Microsoft is now both. Your security team needs to know that Mythos class capability has already found 10,000 critical vulnerabilities in eight weeks of restricted access, and your AI vendor diligence should account for one of those labs in becoming a public company in the next six to twelve months. That's all for this week's Future Bytes news. If you found this useful