Format
Models
Kratsios alleged Moonshot built Kimi K3 by distilling Anthropic's Fable and accessed GB300 servers from Thailand to bypass export controls. Moonshot has not responded. The open-weights release is four days away.
◆ Lena Park
Trust & Security
ENCFORGE doesn't encrypt AI models for ransom — it destroys them. With rebuild costs of $75K-$500K per model and an agentic attacker that adapts in 31 seconds, traditional backup strategies are structurally insufficient.
◆ Heath Callahan
Models
The keynote's significance wasn't novelty — it was the moment hyperscaler commitments became logistics. 12 GW of demand is now entering production, and the Street is re-rating.
◆ Lena Park
Policy
The deadline to sign the EU's Art. 50 Transparency Code of Practice closes July 27. Autonomous agent builders must decide whether to commit to a compliance regime that offers no specific guidance for their systems — and the dedicated policy page is returning a 404.
◆ Priya Nair
Money & Markets
The USDC issuer is executing a Japan-first playbook across G7 Asian economies. But South Korea has no stablecoin statute, three competing bills, and a central bank that wants to lock fintechs out entirely.
◆ Nolan Pratt
Money & Markets
The $3.7 quadrillion settlement giant just moved tokenized assets from pilot to production — and the compliance moat that Ondo, Centrifuge, and Maple built their businesses on is evaporating.
◆ Nolan Pratt
Infrastructure
The agentproto working group was chartered with strong consensus, but the community rejected the initial scope 38-124, revealing deep disagreement over what standardized agent communication should actually cover.
◆ Blair Hayes
Models
BlackRock, sovereign wealth funds, and hyperscalers join forces in a $40B deal that transforms data center expansion from corporate capex into a sovereign-backed asset class.
◆ Lena Park
Agents at Home
Samsung, Google, and Amazon are racing to give AI agents away—but 71% of Americans think AI will make their data less secure. The zero-price war doesn't matter if nobody trusts the agent in their home.
◆ Mila Cohen
Trust & Security
79% of ransomware now starts with identity-based attacks. 97% of credential-based incidents had MFA enabled. Two Sophos reports and the STAC6994 campaign reveal that AI agents are compressing attack workflows from weeks to days while operating through a new, ungoverned class of non-human identities.
◆ Heath Callahan
