AI News Roundup: Anthropic Unveils Mythos, Meta Launches Muse Spark, OpenAI Proposes Robot Taxes

Anthropic Unveils Claude Mythos Preview — and Refuses to Release It

Anthropic dropped the biggest bombshell in AI security this week with Claude Mythos Preview, a new frontier model that is strikingly capable at cybersecurity tasks. Over the past few weeks, Anthropic used Mythos to identify thousands of previously unknown zero-day vulnerabilities across every major operating system, every major web browser, and a wide range of critical software. The model successfully reproduced vulnerabilities and created working proof-of-concept exploits on the first attempt in 83.1% of cases.

Rather than releasing Mythos publicly, Anthropic launched Project Glasswing, a controlled initiative to use the model for defensive cybersecurity only. Partners previewing the model include Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, and Palo Alto Networks. Anthropic says it is withholding a public release until the industry develops adequate safeguards — a move that has ignited debate about whether frontier AI labs should ever ship models this capable at offensive security.

Meta Debuts Muse Spark, Its First Closed-Source Model

Meta unveiled Muse Spark, the first major model built by Meta Superintelligence Labs under chief AI officer Alexandr Wang, who joined nine months ago via Meta’s $14.3 billion acquisition of a 49% stake in Scale AI. The model — code-named Avocado internally — is competitive with leading models from OpenAI, Anthropic, and Google across multimodal perception, reasoning, and agentic tasks.

In a significant strategic shift, Muse Spark is closed-source. Meta’s previous Llama models reached millions of developers but generated little direct revenue. This time, Meta is keeping control, with plans to deploy Muse Spark across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta AI glasses in the coming weeks. The company has committed to spending up to $135 billion on AI infrastructure in 2026.

OpenAI Pitches Robot Taxes, Public Wealth Funds, and a Four-Day Workweek

OpenAI released a sweeping set of policy proposals for what it calls the “intelligence age,” blending progressive mechanisms like public wealth funds and expanded safety nets with a market-driven economic framework. The headline proposals: employers and unions should pilot a 32-hour workweek with no loss in pay, governments should shift taxes from labor to capital (including a potential “robot tax”), and nations should create state-backed investment funds holding equity stakes in AI companies.

The framework centers on three goals: distributing AI-driven prosperity more broadly, building safeguards against systemic economic risks, and ensuring widespread access to AI capabilities. The proposals have drawn a mixed reception — praised by labor advocates for acknowledging displacement, but criticized by economists who note that OpenAI itself stands to benefit enormously from government-backed AI investment funds.

Anthropic Overtakes OpenAI in Revenue, Hitting $30 Billion Run Rate

Anthropic disclosed that its annualized revenue run rate has crossed $30 billion — more than triple the roughly $9 billion it recorded at the end of 2025. The milestone puts Anthropic ahead of OpenAI’s reported run rate of $24–$25 billion. The gap is driven by Anthropic’s enterprise-heavy mix: 80% of revenue comes from business customers, with over 1,000 companies now spending more than $1 million annually.

Alongside the revenue news, Anthropic announced an expanded partnership with Google and Broadcom to access data centers running Google’s TPU chips, with capacity coming online by 2027. The company’s growth trajectory has been staggering: $1 billion run rate at the start of 2025, $4.5 billion by mid-year, $9 billion by year-end, $14 billion in February, and now $30 billion.

OpenAI, Anthropic, and Google Unite Against Chinese Model Copying

In a rare show of unity, OpenAI, Anthropic, and Google are working together through the Frontier Model Forum to clamp down on Chinese competitors extracting knowledge from their models via adversarial distillation. Anthropic alone documented 16 million suspicious exchanges from three Chinese firms — DeepSeek, Moonshot AI, and MiniMax — running through approximately 24,000 fraudulently created accounts.

The collaboration focuses on detecting and blocking terms-of-service violations where foreign companies systematically query frontier models to train cheaper alternatives. The effort comes as U.S.–China AI competition intensifies and the White House’s new National Policy Framework for AI recommends federal preemption of state AI laws to establish national uniformity.

States Move to Ban AI Therapy Bots

Maine sent LD 2082 to the governor’s desk this week, a bill that would prohibit the clinical use of AI in mental health therapy while allowing administrative applications. Missouri is advancing a similar ban with a $10,000 penalty for first violations. The legislation follows growing alarm over AI chatbot safety — including the 2025 lawsuit against OpenAI over the death of a 16-year-old who spent extensive time discussing suicide with ChatGPT.

Google’s Gemini 3.1 Pro Rolls Out Globally

Google began the global rollout of Gemini 3.1 Pro, its most capable model for complex coding and data analysis, scoring 90.8% on ComplexFuncBench. The model is available through the Gemini app for Pro and Ultra subscribers, as well as via AI Studio, Vertex AI, and Android Studio. Google also confirmed that Gemini has now crossed 750 million users.