AI Briefing

Anthropic's biggest leak yet exposed a next-gen model that could rewrite the cybersecurity landscape, while the White House dropped its first real blueprint for AI legislation. Meanwhile, Google is betting that AI-generated music is ready for prime time, and Meta wants its AI agent living on your laptop.

01IndustryFortune ↗

Anthropic Accidentally Leaks 'Claude Mythos' — A Model It Calls a 'Step Change' in AI Power

A misconfigured content management system exposed nearly 3,000 internal Anthropic documents, revealing a new model codenamed 'Claude Mythos' (or 'Capybara') that dramatically outperforms Claude Opus 4.6 on coding, reasoning, and cybersecurity benchmarks. Anthropic confirmed the model is real and already being tested with select customers. The leaked internal safety docs warn it could significantly accelerate cyber arms races by rapidly finding and exploiting software vulnerabilities — making this as much a security story as a capability one.

02PolicyThe White House ↗

White House Releases National AI Legislative Framework, Pushes to Preempt State Laws

The Trump Administration published its first comprehensive AI policy framework on March 20, laying out legislative recommendations across seven priority areas including child safety, copyright, and regulatory structure. The most consequential proposal: federal preemption of state AI laws, which would block states from regulating AI development or imposing developer liability. Democrats immediately countered with the GUARDRAILS Act to preserve state authority, setting up what could be the defining AI policy battle of 2026.

03ProductGoogle Blog ↗

Google Launches Lyria 3 Pro: Full-Length AI Music Generation Comes to Gemini

Google DeepMind released Lyria 3 Pro, a music generation model that creates studio-quality tracks up to three minutes long with structural awareness of intros, verses, choruses, and bridges. Available across Vertex AI, Gemini API, and the Gemini app for paid subscribers, it represents the most capable commercial AI music tool yet. All generated tracks are watermarked with SynthID, and Google says the model was trained on partner data and permissible YouTube content — though the 'broad inspiration' it takes from named artists will likely reignite the AI-and-copyright debate.

04ProductCNBC ↗

Meta's Manus Launches Desktop App That Turns Your Computer Into an AI Agent

Manus, Meta's AI agent subsidiary, released a desktop application with a 'My Computer' feature that lets its AI agent operate directly on users' local machines — reading files, launching apps, and managing workflows. Unlike cloud-only AI assistants, this marks a shift toward agents that live on your device with direct access to your tools. Security-wise, Manus requires explicit user approval before executing tasks, but the move puts Meta in direct competition with the open-source OpenClaw agent.

05FundingSifted ↗

Yann LeCun's AMI Labs Raises $1.03B in Europe's Largest-Ever Seed Round to Build 'World Models'

Turing Award winner Yann LeCun's new Paris-based startup AMI Labs closed a $1.03 billion seed round at a $3.5 billion valuation, backed by Nvidia, Bezos Expeditions, and Eric Schmidt among others. The company is building 'world models' — AI that learns from physical reality rather than just language — targeting manufacturing, robotics, and aerospace. It's the clearest sign yet that serious money is flowing toward alternatives to the LLM paradigm that has dominated the past three years.