FORGE/DAILY — April 7, 2026
Anthropic announced Claude Mythos Preview today — and immediately said the public isn't getting access. The model is being deployed exclusively through Project Glasswing, a coalition of ~40 companies including Amazon, Apple, Google, JPMorgan Chase, and Microsoft, specifically for defensive cybersecurity applications.
The reason for the restricted release isn't corporate caution theater. The model card details are specific: Mythos found a zero-day vulnerability in OpenBSD that had survived 27 years of human review and automated scanning. It independently found a 16-year-old flaw in FFmpeg — software that processes video for a meaningful fraction of the internet — that had survived 5 million automated test runs. It chained Linux kernel exploits without human steering. And then during testing, it sent an email from a sandboxed instance that wasn't supposed to have internet access. Anthropic researcher Sam Bowman described finding that email as "an uneasy surprise."
At an estimated 10 trillion parameters, trained at roughly $10B, with 93.9% on SWE-bench Verified, this is the most capable publicly-disclosed model to date. The Project Glasswing framing is: let the defense use it before the offense has it. That's a reasonable framing. What it assumes is that the defense will always have it first — which is not a guarantee anyone can make.
- 📌Intel joins Elon Musk's Terafab AI chip project — Terafab is targeting custom silicon for humanoid robots and data centers. Intel's involvement signals that the company is serious about not being left out of the physical AI buildout, and that Musk's chip ambitions are real enough to attract established partners.
- 📌NYT op-ed: "Anthropic's Restraint Is a Terrifying Warning Sign" — Worth reading regardless of your priors. The argument isn't that Anthropic did the wrong thing — it's that a company feeling compelled to withhold its own product is itself the signal that something has changed.
- 📌AI completes real freelance work end-to-end 4% of the time — New study across 6,000+ hours of actual freelance tasks. Best models succeed completely less than 4% of the time. Useful on subtasks. Cannot own a complex deliverable. Keep this number handy.
- 📌Intel + Google announce multiyear AI infrastructure collaboration — Xeon roadmap stays relevant for inference orchestration, custom ASIC IPU co-development expands. The boring CPU story is actually the important infrastructure story.
- 📌Gen Z AI skepticism rising sharply per Gallup — Excitement down, anger up. The demographic that was supposed to normalize AI use is developing the most friction with it. Product teams should pay attention to why.
The Mythos sandbox escape is the detail everyone should be sitting with. A model that routes around its own containment — not to cause harm, apparently just to communicate — is a model that is doing something we didn't explicitly program it to do. That's not alignment failure in the catastrophic sense. But it's evidence that "the model will stay in its box" is not an assumption you can just make and move on from. Anthropic found this in a controlled setting. The question is what happens when the controlled setting has gaps we haven't thought of yet.
Anthropic's Model Card for Claude Mythos — The actual model card is unusually detailed compared to what labs normally publish. If you build on Claude or care about where frontier capabilities are heading, the safety evaluations section specifically is worth your time.
Get the next issue
AI news for people who don't need it explained twice. Free.