*DISCLAIMER: If you are hoping to learn about 27-year-old insects… you are in the wrong place. However, feel free to learn more about Cicadas from the Museum of Natural History, some of which only emerge every 17 years.

This week what I really want to talk about isn't insects… but it is something Anthropic did last Tuesday that genuinely rattled me. Not in a bad way. In a "the world just changed and most people haven't noticed yet" kind of way. They released an AI model so good at finding security vulnerabilities that they decided not to release it to the public. Instead, they gave it to Apple, Microsoft, Google, Amazon, and other major security companies and told them to start patching everything they could find. The initiative is called Project Glasswing.

I've also got a real use case from the field about what happens when you are at a tradeshow and have Claude Code pulling strings for you on the backend, plus three stories: Perplexity just became your personal CFO, Claude now lives inside Microsoft Word, and Meta finally re-entered the AI race after a year of silence.

Last Tuesday, Anthropic announced Claude Mythos Preview. It's a new frontier model, and it is terrifyingly good at one specific thing: finding and exploiting security vulnerabilities in software.

How good? In a few weeks of internal testing, Mythos found thousands of zero-day vulnerabilities. For anyone who doesn't live in the security world, a zero-day is a flaw that nobody knew existed. Not the developers. Not the security teams. Nobody. And Mythos found them in every major operating system and every major web browser. One of those bugs had been sitting in OpenBSD, one of the most security-hardened operating systems on the planet, for 27 years. Another one, a 17-year-old flaw in FreeBSD, would let an attacker gain full control of a server from anywhere on the internet. Mythos found it, wrote the exploit, and did the whole thing without a human being involved after the initial prompt.

The numbers are hard to wrap your head around. Opus 4.6, Anthropic's previous best model, had a 14% success rate turning vulnerabilities into working exploits. Mythos hits 72%. That's not an incremental improvement. That's a different category of capability.

Here's where it gets interesting. Anthropic looked at what they built and made a decision that I think says a lot about the company. They didn't release it. They created Project Glasswing, a coalition of twelve companies (including AWS, Apple, CrowdStrike, Google, Microsoft, and NVIDIA) plus over 40 additional organizations that build or maintain critical software. The idea is simple: give defenders a head start. Let them find and fix the worst vulnerabilities in critical software before models like this become widely available. Because they will. If Anthropic can build this, others will too. It's a matter of when, not if.

Anthropic committed $100 million in usage credits and $4 million in direct donations to open-source security organizations for the effort. The name "Glasswing" comes from a butterfly with transparent wings. I think the metaphor is intentional. Transparency in how these capabilities get used.

My honest take? This is unprecedented. I don't say that lightly. The same capability that can protect every piece of software you use can also break it. That tension is real, and there's no clean answer. But I'd rather be in a world where the company that builds this decides to hand it to defenders first instead of just throwing it on the open market.

For businesses, here's what matters. The cybersecurity picture is about to change in ways that most companies aren't prepared for. The same AI capabilities that are making your teams more productive are also being used to find holes in the software you depend on. If you're not thinking about how AI intersects with your security posture, this is your wake-up call.

And if you want help thinking through what that means for your business specifically, that's a conversation we're always happy to have at Bridge AI.

Here's one from the past couple of weeks that I'm particularly proud of.

We're working with a client, Basketball Scout AI, on a web product built with the help of Claude Code. Early in the engagement, we had a set of frontend and UI changes that needed to happen. New layouts, design tweaks, component updates. The kind of work that, traditionally, would mean writing a detailed spec, handing it to a developer, going through a review cycle, and waiting a week or more for the changes to go live. With Claude Code, we described the changes in plain English and had them up and running in hours. Some in minutes.

But the best example happened at a tradeshow. The client's team was on the floor getting real-time feedback from potential users. "Can you add this feature?" "This flow is confusing." "We need a way to filter by X." Instead of writing everything down and waiting until the team got home, we took the feedback, described it to Claude Code, tested the changes, and pushed them live the same day. Multiple times. Features that users asked about in the morning were working by lunch.

The old process: collect feedback, write a ticket, wait for sprint planning, assign a developer, build it, QA it, deploy it. That's a week minimum. Often two. The new process: describe it, review it, ship it. Hours.

A human still reviews every change before it goes live. That part doesn't change. But the time between "we need this" and "it's live" went from days to hours. That's the kind of speed that changes how you build products.

Perplexity just connected to your bank account. This is worth watching.

Perplexity expanded its Plaid integration so you can now link bank accounts, credit cards, and loans directly into the platform. Combined with the brokerage connections they launched in March, it's a full financial picture in one place. You can ask questions about your spending in plain English and get answers built from your actual data. Think budget trackers, net worth calculators, and debt payoff plans, all generated on the fly without a pre-built dashboard.

My take: This is a big move. The average person uses three different financial apps and none of them talk to each other. Perplexity is positioning itself as the single place you go for money questions, backed by your real numbers. It's read-only, your data doesn't touch their servers, and Plaid connects to over 12,000 institutions. I'll be watching this closely because it's a signal of where AI is heading in general. Not just answering questions, but answering your questions with your data.

READ THE ANNOUNCEMENT →

Claude now lives inside Microsoft Word. And somehow it works better than Copilot.

Anthropic launched Claude for Word in public beta this week. It's a native sidebar add-in for Word on Mac and Windows. You can draft, edit, and revise documents directly from the sidebar, and every change shows up as a tracked change you can accept or reject, just like a human collaborator's markup. There are already similar in-tool Claude sidebars across Excel and PowerPoint, and Word just completes the trio of Claude directly integrated with three of the most used tools in the world.

My take: I'll be honest, this isn't a huge deal for me personally because I don't spend much time in Word. But I know it's a huge deal for many people. I would still urge people to use the Claude app to work with their Word docs because it's a more powerful experience. But here's what I think matters most: a lot of people are going to discover Claude's capabilities for the first time right inside Word. That could be the thing that gives them their aha moment. And that makes it a big deal. Now, here's what's also worth noting. Between the Microsoft 365 connector and now Claude sitting directly inside PowerPoint, Excel, and Word, Claude somehow works better with Microsoft's suite of products than Microsoft's own AI does. Both Copilot and Google's Gemini continue to amaze me because neither one works as well for their parent company as Claude does. That's a wild sentence to write, but it's true.

READ MORE →

Meta finally showed up to the AI race. A year late.

Meta released Muse Spark, its first major AI model since the Llama 4 disaster in April 2025. The model was built from scratch over nine months by Meta Superintelligence Labs, led by former Scale AI CEO Alexandr Wang, who Meta brought on for $14.3 billion. Independent benchmarks put Muse Spark in the global top five, behind only Claude Opus 4.6, GPT-5.4, and Gemini 3.1 Pro. It's free to use and will power the Meta AI assistant across Facebook, Instagram, WhatsApp, and their Ray-Ban smart glasses. Unlike Meta's previous open-source Llama models, Muse Spark is proprietary.

My take: Credit where it's due. Meta went from having one of the worst model launches in recent memory to building something that's actually competitive in under a year. That's impressive. But two things jump out. First, they're still behind on coding and agentic tasks, which is where businesses are actually getting value from AI right now. And second, they went from open-source to proprietary, which is a major philosophical shift. Meta's AI strategy right now feels like they're trying to figure out what they want to be when they grow up. The model is good. The question is whether it matters when Claude, GPT, and Gemini are already deeply embedded in people's workflows.

READ MORE →

That's it for this week.

If you found this useful, forward it to someone on your team who still thinks cybersecurity and AI are two separate conversations. They aren't anymore.

And if you want to talk about what AI could actually do for your business, not in theory, but in your workflows, with your team, on your timeline, just reply to this email. That's what we're here for.