Anthropic's Mythos model found vulnerabilities in every major OS, plus a 754B open-weight coding agent and Intel joins Terafab

Picture this: an AI scans every major operating system and web browser in existence - and finds a security flaw in all of them. No human intervention. That's not a thought experiment. That's what Anthropic's new Mythos model reportedly did, and it's just the opening act of a much bigger initiative announced yesterday.

There's a lot happening today - a massive open-weight agent model, Intel betting on a controversial chip factory, and a quiet but genuinely useful offline dictation app from Google. Let's get into it.

On which benchmark did Anthropic's new Mythos model specifically demonstrate its agentic cybersecurity capabilities - helping find vulnerabilities across operating systems and browsers?

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

Anthropic launched a preview of its new Claude Mythos model yesterday as part of Project Glasswing - a cybersecurity coalition that includes Apple, Google, Amazon Web Services, Nvidia, Microsoft, and more than 45 other organizations. The goal is automated, large-scale vulnerability detection with virtually no human intervention required.

The headline finding is striking: the model identified security problems in every major operating system and web browser during testing. Launch partners get early access to Mythos Preview specifically for defensive security work - scanning their own systems for weaknesses before attackers find them.

This is meaningfully different from a chatbot that helps write security reports. Mythos is being positioned as an active, agentic scanner - one that can work through complex codebases and system configurations autonomously. The fact that governments could eventually be customers adds another layer to how significant this initiative could become.

Read the full story →

While everyone watches the big labs, Z.AI quietly shipped something remarkable: GLM-5.1 - an open-weight 754 billion parameter model built specifically for agentic engineering tasks. It's not optimized for clean single-turn benchmarks. It's built to actually work.

The standout claim: GLM-5.1 can sustain autonomous execution for up to 8 hours - handling long-horizon tasks that would cause most models to lose context, repeat themselves, or go off-rails. It also achieves state-of-the-art performance on SWE-Bench Pro, the toughest version of the software engineering benchmark, outpacing its predecessor by a significant margin.

Because the weights are open, developers can run this locally - important for teams with data privacy requirements or API cost concerns. If you've been experimenting with coding agents, this one is worth a serious look.

Read the full story →

Intel confirmed yesterday it's joining the Terafab project - Elon Musk's plan to build a new AI chip factory in Austin, Texas. The facility would supply chips to SpaceX (now merged with xAI) and Tesla as Musk pushes toward his 'robot army' vision spanning self-driving cars, humanoid robots, and AI inference at scale.

Intel's exact scope of contributions hasn't been fully disclosed, but the involvement of a major U.S. chipmaker signals this is more than a vanity project. Building domestic semiconductor capacity has become a national priority, and Terafab - if it delivers - would represent a significant new node in U.S. AI infrastructure.

The timing is notable given Intel's ongoing strategic reshaping. Signing onto a high-profile domestic AI chip project gives Intel a clearer narrative around its future in the semiconductor space - even if the details remain fuzzy.

Read the full story →

Google quietly released a new AI-powered dictation app for iOS this week - and the key feature is that it works completely offline. It runs on Gemma models on-device, putting it directly in competition with apps like Wispr Flow that have built real followings among power users.

The offline-first approach isn't just a technical achievement - it's a privacy pitch. Everything processes locally, which matters for anyone dictating sensitive notes, legal documents, or anything they'd rather not route through a cloud server. For frequent travelers or anyone in spotty-connectivity situations, it's an immediate practical win.

If you're building a voice-first workflow and haven't locked in a dictation solution, this is a good moment to compare options. And while you're thinking about your AI toolkit - 60sec.site is worth a look if you need to spin up a website fast using AI. No coding required, live in under a minute.

Read the full story →

The friction between AI music generation and the traditional music industry is getting more specific - and more contentious. According to a Financial Times report, Suno is struggling to reach licensing deals with Universal Music Group and Sony Music Entertainment. The sticking point isn't just money. It's about whether users can share their AI-generated tracks outside the Suno platform.

Universal's position is reportedly that AI-generated tracks should stay inside apps like Suno and not spread freely across the internet. That's a fundamental restriction on what users can actually do with the music they create - and it would dramatically limit the appeal of these tools for anyone who wants to publish, post, or monetize their output.

Suno's whole value proposition depends on users being able to use what they make. If labels get their way on containment, the business model gets a lot harder to sustain. This one will have implications for every AI music platform trying to go legitimate.

Read the full story →

The answer is CyberSecEval! Anthropic specifically highlighted Mythos's performance on cybersecurity-focused evaluations as part of the Project Glasswing announcement - demonstrating the model's ability to autonomously identify vulnerabilities across operating systems and browsers in a way that goes well beyond what general-purpose benchmarks like MMLU or HumanEval measure.

AI scanning every major OS and browser for vulnerabilities with no human in the loop - does that feel like a massive security win to you, or does it make you a little nervous? I'm genuinely curious where people land on this one. Hit reply and let me know - I read every response!

That's it for today - a lot packed into one edition. If you want to dig into any of these topics further, head over to dailyinference.com where we cover AI news every single day. See you tomorrow! 👋