⇤ BACK

Weekly AI Roundup: Anthropic Built a Hacker, Meta Built a Lab, and Everyone Built a Wall Around China

· By · clzd.me

Anthropic Made a Model That Hacks Things — Then Refused to Ship It

Anthropic announced Claude Mythos Preview this week. It finds and exploits zero-day vulnerabilities across every major OS and browser. Cool. Also terrifying. And no, you can’t have it.

Over 50 orgs are getting access through Project Glasswing with $100M+ in usage credits. The pitch is defensive — let security teams find holes before attackers do. The subtext is that AI just got better at breaking software than most humans are. Anthropic is trying to control the rollout, which is respectable, but let’s be honest: if one lab built this, others aren’t far behind.

Also this week, Anthropic shipped Claude Managed Agents into public beta. Composable APIs for cloud-hosted agent workflows. Less sexy than a model that pops zero-days, but this is the product that turns Claude from a chatbot into infrastructure. Worth watching.

Meta Finally Shipped Something

Muse Spark is the first model from Meta Superintelligence Labs — the unit they built around Alexandr Wang after that $14.3B Scale AI deal. It’s a general-purpose model. Benchmarks aren’t earth-shattering. Nobody’s switching from Claude or GPT because of it.

But that’s not the point. The point is Meta stopped treating AI as a recommendation engine for Instagram Reels and started building foundation models with serious talent and serious money. First models from new labs are never the ones that matter — it’s the second and third that’ll tell you if the investment was real. Meta has something nobody else does though: a billion users ready to consume whatever they build next.

OpenAI, Anthropic, and Google Agree on Exactly One Thing: China

Three companies that spend most of their time trying to kill each other just agreed to share intelligence through the Frontier Model Forum. That alone tells you how bad the distillation problem is.

Anthropic documented 16 million adversarial distillation attempts from three Chinese companies running through 24,000 fake accounts. That’s not scraping — that’s a factory operation. They bombard the API with prompts engineered to map internal model behavior, then train cheaper clones on the outputs. It’s industrial-scale IP theft dressed up as API usage.

This is the first time the Forum has produced anything beyond press releases. Expect API rate limiting to get more aggressive and expect this to become a policy football fast. The AI cold war just found its first real coalition.

Also Noted

  • Musk filed to remove Altman and Brockman from OpenAI. Still trying to restore the non-profit structure. This lawsuit will outlive us all.
  • Broadcom expanded chip deals with Google and Anthropic — 3.5 GW of TPU compute for Anthropic alone. The hardware layer is where the real war is.
  • Chrome zero-day actively exploited (CVE-2026-5281) — use-after-free in Dawn, RCE via crafted HTML. Patch your browser. Today. Not tomorrow.
  • Claude had back-to-back outages on April 7-8. Users hitting errors on Sonnet 4.6. Ironic timing given the Managed Agents launch.

Dispatch is the weekly AI roundup at clzd.me — snarky, sourced, and probably running on the same models it’s writing about.