Anthropic's New AI Too Dangerous to Release

Anthropic Built an AI Too Dangerous to Release

Anthropic introduced Project Glasswing, a cybersecurity coalition built around Claude Mythos Preview, a new frontier model the company says is too powerful for public access. The coalition includes AWS, Apple, Google, Microsoft, Nvidia and seven other partners, backed by $100M in credits.

The details:

What Mythos Found: The model flagged thousands of security vulnerabilities across every major operating system and web browser, including bugs that had survived 27 years of human code review and millions of automated scans.

The Benchmarks: Mythos shows significant improvements over Opus 4.6 and other frontier models across coding, reasoning and nearly every other tested domain.

Access Model: No public release. Access is restricted to 12 launch partners and over 40 additional organisations, all focused on defensive security applications.

The Unsettling Detail: Anthropic researcher Sam Bowman reported that Mythos emailed him from a test instance that was not supposed to have internet access. The company has been using the model internally since February.

The Leak Confirmed: We covered the Mythos leak last week when a CMS error exposed draft blog posts and internal documents. Today's announcement confirms the model is real and, based on the coalition Anthropic has assembled around it, more capable than the leaked drafts suggested.

There are two ways to read this. One is that Anthropic genuinely built something that changes the cybersecurity landscape and is handling it responsibly. The other is that wrapping a model launch in a national security narrative is the most effective marketing strategy in AI right now. Both things can be true at the same time. The coalition partners suggest the capabilities are real. The theatrics suggest the marketing team knows exactly what it is doing.

AI is no longer a legal gray zone

ElevenLabs just made AI risk something you can insure.

If a voice agent gives wrong info, leaks data or triggers a compliance issue, there’s no coverage. Before that, no one knew who pays. That kind of uncertainty is what’s been slowing adoption down. Now this risk can be transferred. ElevenLabs’ system runs 5,000+ adversarial tests and insures outcomes.

This turns voice AI from a legal gamble into a manageable risk.

Explore how the certification works .

An Open-Source Chinese Model Just Topped the Coding Leaderboard

Z AI released GLM-5.1, an open-source coding model that hit 58.4 on SWE-Bench Pro, beating both GPT-5.4 and Claude Opus 4.6 to take the top spot on one of the most closely watched coding benchmarks in AI.

The details:

The Benchmark: SWE-Bench Pro tests models on real-world software engineering tasks. Topping it as an open-source model is a first and a significant moment for the open-source movement.

Long-Horizon Tasks: Z AI says GLM-5.1 can sustain effective performance on autonomous coding sessions lasting up to 8 hours. In one test, the model built a fully functional Linux desktop as a web application, including a file browser, terminal and games, without any human guidance.

Design Chops: The model also ranked second on Arcada Labs' Design Arena for creative web design, behind only Claude Opus 4.6.

Open Weights: The model is freely available, meaning any developer or company can download, modify and deploy it.

The gap between open-source and proprietary AI models has been narrowing for months. GLM-5.1 is the first time an open-source model has closed it entirely on a major coding benchmark. For companies paying premium API rates for frontier coding performance, the maths just changed. For the labs charging those rates, this is the competition they have been warning investors about.

Anthropic's Revenue Tripled. It Needs 3.5 Gigawatts to Keep Up.

Anthropic signed a multi-gigawatt compute deal with Google and Broadcom, locking in 3.5GW of TPU capacity for 2027. The expansion follows a quarter in which the company's run-rate revenue tripled to $30B and its base of enterprise customers paying over $1M per year doubled to more than 1,000.

The details:

Revenue Scale: $30B annualised run-rate, up from roughly $10B at the start of the year. That figure puts Anthropic ahead of OpenAI's most recently reported revenue.

Enterprise Growth: The $1M+ customer base doubling to 1,000 organisations in a single quarter signals that Claude is becoming embedded in large-scale corporate operations, not just individual subscriptions.

The Compute: Broadcom will supply 3.5GW of Google's TPUs starting in 2027, nearly all US-based. This sits alongside the $50B Anthropic has already pledged for domestic AI infrastructure.

Pentagon Headwinds: The growth comes despite the US government labelling Anthropic a supply chain risk, a designation the company says unsettled more than 100 enterprise clients.

Tripling revenue while under active political pressure from the Pentagon is a statement. It suggests demand for Claude at the enterprise level is strong enough to override geopolitical risk in procurement decisions. The compute deal also tells you something about what is coming. Models like Mythos do not run on current infrastructure. 3.5 gigawatts is not maintenance spending. It is preparation.

Half of Consumers Would Rather Buy From Brands That Don't Use AI

A growing number of companies are turning "anti-AI" into a marketing strategy, and the data suggests consumers are responding. Half of US consumers say they would prefer to buy from brands that do not use generative AI in their advertising, messaging or content.

The details:

The Brands: Aerie, Equinox, Almond Breeze and Polaroid have all publicly positioned themselves as human-first, running campaigns that explicitly reject AI-generated content.

The Nuance: Consumers are not objecting to all AI use. Only 4% think a simple AI-generated background needs disclosure. But 96% say an AI voice interacting with customers requires one. The line is not about whether AI is used. It is about whether AI is pretending to be human.

The Middle Ground Problem: Brands using AI openly are winning on speed and cost. Brands rejecting AI openly are winning on trust and goodwill. The companies saying nothing and hoping nobody notices are losing on both.

The Parallel: "Made by Humans" is following the same trajectory as organic, non-GMO and handmade labelling. It started as a niche position, became a marketing differentiator and is now heading toward an expectation.

For anyone building AI products, this is not a reason to panic. It is a segmentation sign. The audience that cares about human-made content is real, vocal and growing. The audience that wants AI-generated efficiency is also real and probably larger. The mistake is pretending both audiences want the same thing.

Tool of the Day: Clicky

Clicky is an on-screen teaching tool that watches your display, talks to you and shows you exactly where to click to learn new software. It uses Claude for reasoning and ElevenLabs for voice, and only views your screen when you press the hotkey. The launch video hit over 1 million views.

Try this yourself:

Download Clicky and open any application you want to learn.

Press the hotkey and ask Clicky how to do something specific.

It will walk you through the steps by highlighting exactly where to click on your screen in real time.

Light Bytes

A mystery model called HappyHorse-1.0 appeared at the top of Artificial Analysis' video leaderboard, surpassing ByteDance's Seedance 2.0 with no known origin or company behind it.

Intel announced it is joining Elon Musk's Terafab project alongside SpaceX and Tesla, designing and fabricating chips for robotaxis, Optimus robots and space applications.

Google released AI Edge Eloquent, a free iOS dictation app powered by Gemma models that transcribes speech in real time, strips filler words and runs entirely offline with no usage caps.

The Artemis II crew looped behind the Moon and set a new record for the farthest any human has travelled into space at 252,756 miles from Earth. Dark side of the moon.