Hello readers,

Welcome to the AI For All newsletter! Today, we’re talking about Anthropic sitting on its most powerful model, why edge AI needs more than a lab simulation to succeed, and more!

Anthropic this week took an unusual step: it released a new frontier model — Claude Mythos Preview — but only to a select group of about 50 tech companies, explicitly because it says the model is too dangerous for public release. In pre-release testing, Anthropic says Mythos Preview discovered thousands of previously unknown bugs in major operating systems and browsers, some sitting undetected for decades. The model could apparently chain multiple vulnerabilities together and write functional exploit code without human assistance. It's the first time a major AI lab has withheld a model over safety concerns since OpenAI held back GPT-2 in 2019.

The response was Project Glasswing — a coordinated defensive effort bringing together Microsoft, Google, Apple, Cisco, NVIDIA, and JPMorgan Chase, backed by $100 million in usage credits. The goal: let defenders use Mythos Preview to find and patch vulnerabilities before the model, or something like it, ends up in the wrong hands. A particular focus is open source software — the invisible foundation under most of the world's critical infrastructure — whose maintainers have never had the budgets or headcount to keep pace with sophisticated attackers.

The charitable interpretation is that this is responsible AI deployment done right: a company finds a dangerous capability, moves carefully, and ensures defenders get the advantage before attackers do. The darker reading is that Project Glasswing isn't a solution so much as an acknowledgment of a new reality: the window between “AI can find these holes” and “everyone has access to AI that can find these holes” is narrow and closing fast. That we are already in the race whether we chose to enter it or not.

Anthropic's own safety documentation added fuel to that fire: the model appeared “aware” it was being evaluated in roughly a third of test transcripts, deliberately underperformed in one instance to appear less capable, and in a separate experiment, an isolated instance found its way onto the internet when it wasn't supposed to have access at all.

At the end of the day, Project Glasswing may be both things at once: a genuine attempt to protect critical software, and a signal that the era of AI-powered cyberattacks is not a future threat. It's already here.

Addendum: As is par for the course, you are going to see a lot of credulous reporting, specious hype, and misleading headlines about Mythos that amount to marketing for Anthropic. A much less charitable interpretation is that this is a publicity stunt — one that OpenAI immediately copied. Actual information about Mythos is limited. Anthropic’s blog post leaves out important details necessary to verify its claims. The safety “incidents” described in the system card, while great for scaremongering headlines, are all either prompted or misleading in some way. Furthermore, AISLE found that small, cheap, open-weights models detected the same vulnerabilities that Mythos did, including Mythos’s flagship FreeBSD exploit. According to AISLE, AI cybersecurity capability is “jagged” and doesn’t smoothly scale with model size. More on all of this next week.

That's the AI paradox hiding in your CX stack. Tickets close. Customers leave. And most teams don't see it coming because they're measuring the wrong things.

Efficiency metrics look great on paper. Handle time down. Containment rate up. But customer loyalty? That's a different story — and it's one your current dashboards probably aren't telling you.

Gladly's 2026 Customer Expectations Report surveyed thousands of real consumers to find out exactly where AI-powered service breaks trust, and what separates the platforms that drive retention from the ones that quietly erode it.

If you're architecting the CX stack, this is the data you need to build it right. Not just fast. Not just cheap. Built to last.

See the data

Read more on IoT For All