AI News Roundup: Anthropic Delays Claude Mythos Over Cyber Risks, Meta Launches Muse Spark, Big Three Unite Against Model Theft

Anthropic Delays Claude Mythos After Model Discovers Zero-Day Exploits

Anthropic made waves this week by announcing it would not release its most powerful AI model, Claude Mythos, to the general public. Disclosed on April 7–8 and formally restricted on April 12, the decision came after internal testing revealed Mythos could autonomously discover and exploit zero-day vulnerabilities in every major operating system and web browser.

Rather than shelve the model entirely, Anthropic is restricting access through a new initiative called Project Glasswing, a consortium of technology companies that will use the model under strict oversight. The European Commission publicly backed the staged rollout, citing large-scale cyber risk concerns. The move positions Anthropic as a “responsible leader” ahead of its expected $60 billion+ IPO later this year, though some analysts warn the delay could reduce near-term API revenue.

Meta Launches Muse Spark — Its First Proprietary AI Model

Meta debuted Muse Spark on April 8, its first major AI model built from scratch by Meta Superintelligence Labs under new Chief AI Officer Alexandr Wang. The launch marks a dramatic shift from Meta’s open-source Llama strategy: Muse Spark is proprietary, though the company says it hopes to “open-source future versions.”

The results are striking. Muse Spark scored 52 on the Artificial Analysis Intelligence Index, nearly triple Llama 4 Maverick’s score of 18, while using over an order of magnitude less compute. Wang, the 29-year-old former Scale AI co-founder, was recruited by Mark Zuckerberg last summer as part of a total overhaul of Meta’s AI operations. This is the first proof point that the bet is paying off.

OpenAI, Anthropic, and Google Unite Against Chinese Model Copying

In a rare display of cooperation, OpenAI, Anthropic, and Google announced they are sharing intelligence through the Frontier Model Forum to combat unauthorized model copying by Chinese AI companies. The alliance targets so-called “adversarial distillation” — where competitors systematically query frontier models to extract and replicate their capabilities.

Anthropic documented 16 million exchanges from three Chinese companies — DeepSeek, Moonshot AI, and MiniMax — running through approximately 24,000 fraudulently created accounts. The Frontier Model Forum, originally founded by the three companies and Microsoft in 2023, is now pivoting to active intelligence-sharing and enforcement to protect intellectual property across the industry.

PwC Study: 20% of Companies Capturing Three-Quarters of AI’s Economic Value

A new PwC AI Performance study released today finds that a small group of companies is pulling sharply ahead in the race to generate real financial returns from AI. The top performers are 2.6x more likely to report that AI improves their ability to reinvent their business model, and they treat the technology as a growth engine rather than just a cost-cutting tool.

The key differentiator? Leaders invest in the foundations that make AI scalable and reliable — data infrastructure, governance, and cross-functional integration — while laggards continue to treat AI as siloed experiments. The finding echoes a growing consensus that the AI divide between companies is widening, not narrowing.

Stanford AI Index 2026: Generative AI Adoption Outpaces the Internet

Stanford HAI’s annual AI Index report dropped with a landmark finding: generative AI reached 53% population adoption within just three years, faster than the personal computer or the internet. The estimated annual value of generative AI tools to U.S. consumers has reached $172 billion, with the median value per user tripling between 2025 and 2026.

But the report also flagged concerns. The Foundation Model Transparency Index — which tracks how openly companies disclose training data, compute, and risks — saw average scores drop from 58 to 40 points. And regulatory fragmentation is accelerating: 47 countries now have active AI legislation, but only 12 have established enforcement mechanisms. There were 156 documented enforcement actions globally in 2025, up from just 43 in 2024.

States Push Forward on AI Legislation

State lawmakers have introduced over 600 AI bills in the 2026 legislative sessions so far. This week, Nebraska passed a bill regulating minors’ interactions with conversational AI, Maryland passed an AI-driven pricing bill, and Maine prohibited unlicensed AI therapy services. Indiana, Utah, and Washington enacted laws barring health insurers from using AI as a sole basis for denying or modifying claims.