AI Briefing

Today's throughline is who gets to build, ship, and switch off the most powerful AI. Jeff Bezos finally said out loud what his secretive $41B startup is for, Washington yanked Anthropic's flagship model off the grid worldwide, and a self-spreading worm turned everyday AI coding assistants into credential thieves inside Microsoft's own repos. Meanwhile a Chinese lab dropped a trillion-parameter coder for free and DeepMind quietly published its roadmap for what comes after human-level AI.

01FundingeWeek ↗

Jeff Bezos Breaks Cover: Prometheus Raises $12B to Build an 'Artificial General Engineer'

Bezos and former Google executive Vik Bajaj revealed that their year-old startup Prometheus raised $12 billion in a Series B, vaulting it to a $41 billion valuation backed by JPMorgan, Goldman Sachs and BlackRock. Rather than chasing AGI chatbots, Prometheus wants an 'artificial general engineer' that compresses the design-and-manufacture cycle for physical products like jet engines and drug compounds by 10x or more. It matters because it points the AI capital wave at the physical economy, and a huge chunk of the raise is going straight into compute because the manufacturing training data simply doesn't exist on the internet.

02PolicyNextgov/FCW ↗

Washington Pulls the Plug: U.S. Forces Anthropic to Disable Claude Fable 5 and Mythos 5 Worldwide

Three days after launching Fable 5 as its most capable public model, Anthropic was hit with a Commerce Department export-control order barring access by any foreign national, citing national security. Unable to verify citizenship in real time, the company disabled both Fable 5 and Mythos 5 for every customer on Earth and began issuing refunds. This is the first time the U.S. has used export-control authority to switch off a commercial AI model globally, treating a frontier model less like software and more like a controlled weapon, and the irony is sharp given Anthropic spent years asking governments for exactly this kind of power.

03IndustryThe Hacker News ↗

Miasma Worm Hijacks 73 Microsoft GitHub Repos by Weaponizing AI Coding Assistants

A self-replicating supply-chain worm called Miasma compromised 73 Microsoft repositories across the Azure, Azure-Samples, Microsoft and MicrosoftDocs GitHub orgs. The clever part: instead of poisoning a package, the malicious commit planted config files (.claude, .gemini, .cursor, .vscode) that auto-execute a credential-stealing payload the moment a developer opens the repo in Claude Code, Gemini CLI, Cursor or VS Code. It then uses harvested AWS, Azure, GCP and GitHub tokens to copy itself into every repo the victim can write to. GitHub disabled the affected repos, but the attack exposes how the trust model behind AI-assisted development can be turned into a propagation engine.

04Open SourceCrypto Briefing ↗

Moonshot AI Drops Kimi K2.7-Code, a Free 1-Trillion-Parameter Coding Model

Beijing-based Moonshot AI released Kimi K2.7-Code, a Mixture-of-Experts model with 1 trillion total parameters (32B active) under a permissive Modified MIT license on Hugging Face. It posts double-digit benchmark gains over its predecessor, a 21.8% jump on Kimi Code Bench v2 and 31.5% on multi-language tests, while cutting reasoning-token usage by 30% to fight 'overthinking.' It's the company's fifth major release in under a year and another sign that open-weight models are closing the gap with closed frontier systems, this time aimed squarely at the agentic coding workloads where token costs bite hardest.

05ResearchGoogle DeepMind ↗

DeepMind Maps the Road Past Human-Level AI in 'From AGI to ASI'

Google DeepMind published a roadmap, authored by AGI pioneers Shane Legg and Marcus Hutter among others, examining how AI might keep advancing once it matches human intelligence. It defines artificial superintelligence as a system more capable than large coordinated teams of human experts, then lays out four possible pathways there: scaling current models, new algorithmic paradigm shifts, recursive self-improvement, and superintelligence emerging from large multi-agent collectives. The paper argues the popular image of a single AGI 'step change' may be wrong, and that we should instead prepare for a series of transformative breakthroughs across science and technology.