Mythos 5 access widened as global AI regulation expands

Anthropic / Claude ecosystem

No significant new developments.

Frontier model providers

OpenAI Launches Codex in ChatGPT Mobile App for General Availability

Codex is now generally available in the ChatGPT mobile app, bringing full AI coding assistance to iOS and Android with cross-device continuity and security features.

AI developer tooling & infrastructure

Stably Ships Orca, Open-Source IDE for Fleets of Coding Agents

Orca is a new IDE that orchestrates multiple coding agents in parallel using isolated git worktrees, aiming to simplify fleet-based software development.

New Enterprise-Ready MCP Specification Introduces Security Challenges

A new version of the Model Context Protocol designed for enterprises makes AI agent communication cloud-native but also raises new security concerns, according to an expert analysis.

Cloud & platform providers

NVIDIA Expands AWS AI Infrastructure with New Blackwell GPU Instances

AWS introduced new GPU instances powered by NVIDIA RTX PRO 4500 Blackwell GPUs, delivering accelerated inference and vector indexing for AI workloads.

AI policy, regulation & governance

US Allows Anthropic to Release Mythos 5 to Trusted Partners

The US government authorized Anthropic to deploy its Mythos 5 model to over 100 trusted US organizations and agencies for cybersecurity use, partially lifting the earlier export ban. This marks a significant shift in access to the frontier model.

EU Considers Designating AWS and Azure as Gatekeepers Under DMA

The European Commission may extend the Digital Markets Act to cloud infrastructure providers like AWS and Azure, imposing gatekeeper obligations on the dominant platforms.

Australia Plans Tougher Online Safety Laws with AI Focus

The Australian government plans to introduce a broader digital duty of care and strengthen the eSafety Commissioner's powers, targeting AI-generated harmful content and social media enforcement.

Senator David Pocock Urges PM to Stop Australian Content Used in AI Training

Independent Senator David Pocock called on the Prime Minister to halt tech companies from using Australian content for AI model training without consent, escalating the AI copyright debate.

Industry & market moves

Oasis Founder Dawn Song Joins Meta Superintelligence Labs as VP of AI Research

Noted computer security and AI safety researcher Dawn Song has taken a senior role at Meta's Superintelligence Labs, bringing expertise in privacy and secure AI.

Quantifind Raises $200M Growth Investment for AI-Native Risk Intelligence

Quantifind secured $200 million from Summit Partners to advance its AI-driven risk intelligence and agentic middleware for financial crime compliance.

AI Startup General Intuition Raises $320M at $2.3B Valuation

General Intuition raised $320 million led by Khosla Ventures and General Catalyst, using video game environments to train AI agents. The round values the company at $2.3 billion.

AI product & feature launches

Toss Partners with Poseidon to Bring 30 Million Users into AI Data Economy

Korean fintech Toss will let its 30 million users contribute real-world data for AI training and get paid, a first-of-its-kind model in Korea, with plans to expand globally.

MiniMax Launches General-Purpose AI Agent for Long-Horizon Tasks

MiniMax Agent is a new general intelligent agent capable of long-horizon planning and multi-subtask execution with multimodal outputs, emphasizing requirement-driven development.

Open-Source Coding Model Ornith-1.0 Learns Its Own Training Scaffold via RL

DeepReinforce released Ornith-1.0, an open-source coding model family that can write its own reinforcement learning training scaffold, enabling self-improving code generation under MIT license.

Research with immediate practical relevance

Google Accelerates Gemini Nano on Pixel with Frozen Multi-Token Prediction

Google revealed a technique to retrofit Multi-Token Prediction onto frozen Gemini Nano models, significantly improving on-device AI speed and energy efficiency for Pixel devices.

NVIDIA Research: AI Model Self-Corrects Its Training Metric Mid-Run

In an autonomous training run, a 30B NVIDIA Nemotron model fixed a broken evaluation metric on its own, improving external benchmarks. The A-Evolve experiment shows self-correcting AI development.