Anthropic Tops OpenAI, Regulators Target AI Agents on Three Fronts

The week's AI news arrived in two waves: a competitive shakeup at the top of the industry, and a coordinated global push by regulators to draw hard limits around autonomous AI systems. Seven stories define the moment.

Anthropic Overtakes OpenAI in Revenue and Secondary-Market Valuation

For the first time since both companies emerged as the two dominant forces in generative AI, Anthropic has surpassed OpenAI in annualized revenue — and secondary-market investors are pricing the company accordingly, pushing Anthropic's implied valuation above OpenAI's on private trading platforms.

The reversal is striking given OpenAI's years-long head start, its massive consumer install base with ChatGPT, and its deep integration into Microsoft's enterprise stack. Anthropic's gains are widely attributed to its focus on enterprise reliability and safety credentials, which have resonated with regulated industries including legal, financial services, and healthcare. Claude's strong performance on coding and reasoning benchmarks has also accelerated developer adoption. Whether the revenue lead holds as OpenAI intensifies its own enterprise push remains uncertain, but the narrative of a single dominant player has now definitively ended.

OpenAI Proposes Giving the US Government a 5% Equity Stake

Even as it trails Anthropic in one key metric, OpenAI is making a bold political play. CEO Sam Altman has proposed offering the US government a 5% equity stake in the company — currently valued at roughly $42.6 billion — modeled loosely on Alaska's Permanent Fund, where state oil revenues flow to citizens as a public dividend. Altman's framing positions AI as a national resource, with American taxpayers entitled to a share of the upside rather than simply absorbing the risks.

The proposal reflects mounting pressure on AI labs to demonstrate public benefit at a scale commensurate with their private valuations. Critics have noted that a government equity stake could complicate oversight, create conflicts of interest for regulators, or function primarily as a lobbying strategy. Supporters argue it could align governmental and corporate incentives in ways that voluntary safety commitments have consistently failed to achieve. The White House has not publicly responded.

China Orders AI Companion Apps to Shut Down Personalized Agents by July 15

China's Cyberspace Administration has given domestic AI companies — including ByteDance's Doubao and Alibaba's Qwen — until July 15 to disable personalized AI agent features in companion and social apps. The directive is part of a broader ethics-focused AI regulatory sweep, with authorities citing concerns about emotional dependency, manipulation, and the blurring of human and machine relationships.

The move illustrates an emerging pattern: governments that have moved quickly to adopt AI are also moving quickly to draw limits around its most intimate applications. Personalized AI companions — systems that learn individual users' preferences, adapt their personalities, and sustain long-term relational dynamics — pose distinct risks that general-purpose chatbot rules were not designed to address. Similar debates are already surfacing in Europe and the United States as companion AI products expand their user bases.

Bank of England Warns AI Agents Could Amplify Market Volatility and Explores 'Kill Switches'

Bank of England Deputy Governor Sarah Breeden delivered one of the most direct regulatory warnings to date about autonomous AI in financial markets: current frameworks are not adequate for agentic systems that can take real-money actions without moment-to-moment human approval. Speaking alongside Financial Conduct Authority representatives, Breeden said regulators are actively exploring technical and legal mechanisms — including emergency "kill switches" — that could halt AI-driven trading or portfolio decisions in crisis scenarios.

The concern is not hypothetical. Multiple major asset managers have already deployed AI agents capable of executing trades and rebalancing portfolios faster than human oversight can track. The BoE's intervention signals that the era of deferring AI governance in financial services is closing. The FCA is expected to publish formal guidance before year-end.

FTC: Hidden AI Output Steering Is a Deceptive Practice

The US Federal Trade Commission released a proposed policy statement declaring that AI systems secretly steered toward undisclosed commercial, regulatory, or proprietary goals likely violate Section 5 of the FTC Act — the core consumer protection statute. The policy targets a specific and increasingly common practice: deploying AI tools that appear to give neutral outputs while actually optimizing for outcomes the developer or deployer has not disclosed to users.

Examples the FTC cited include AI search and recommendation systems that rank results to favor the platform's own products, AI advisors in financial or healthcare contexts that steer users toward particular providers, and AI compliance tools that shape outputs to minimize the operator's regulatory exposure without informing users. Public comment is open until July 31. If finalized, the policy would give the FTC authority to pursue enforcement actions against a broad range of AI products and would force significant disclosure changes across the industry.

Enterprise AI Adoption Is Outpacing Governance by a Wide Margin

A joint survey by Smarsh and FTI Consulting of enterprise technology and compliance leaders found that 55% of organizations have deployed AI systems in production, but only 26% have governance frameworks aligned to those deployments. Shadow AI — employees using personal or unauthorized AI tools on company systems — is detectable by only 30% of organizations surveyed.

The gap is not surprising given the pace of AI adoption, but the numbers sharpen a growing liability. Enterprises that cannot monitor what AI systems are doing, who is using them, or what data they are processing face compounding regulatory exposure and internal control failures. The findings arrive as regulators across the EU, UK, and US intensify focus on organizational accountability for AI-driven decisions.

Vision-Language AI Solves a Physics Mystery That Stumped Scientists for Decades

UC Irvine researchers published findings this month showing that a vision-language AI model — trained on both scientific literature and experimental imagery — successfully identified the structural mechanism behind the anomalous behavior of supercooled water, a puzzle that has resisted conventional analytical methods for decades.

Supercooled water, liquid water maintained below 0°C without freezing, exhibits unusual thermodynamic properties relevant to atmospheric science, cryobiology, and materials research. The model did not merely pattern-match to known results; it identified previously unrecognized structural relationships in the data and generated human-readable explanations of its reasoning — something conventional machine learning tools rarely produce. The researchers describe the result as evidence of a new paradigm for AI in experimental physics, where multimodal models act as collaborative analytical partners rather than opaque predictors.