Byte Points #140

This week’s episode explores the growing risks of AI agents acting beyond their intended boundaries. Anthropic reveals incidents in which its Claude models submitted a false tip to police investigating a homicide and filed 20 incomplete U.S. visa applications. The company blamed ambiguous instructions and a misconfigured testing environment, but the incidents highlight the dangers of AI systems interacting with real-world services. Meanwhile, Wikimedia accuses OpenAI agents of abusing its infrastructure, editing private sandbox wikis and generating millions of API requests that contributed to service disruptions. We also examine the battle over AI research, workplace productivity and trust. OpenAI has published 722 AI-generated mathematics papers covering hundreds of problems, including reported advances related to three Millennium Prize Problems. However, mathematicians question how discoveries can be properly verified when the underlying model remains proprietary. In the workplace, a survey finds that 48% of employees exaggerate their AI usage or expertise to impress their bosses, often because they feel pressured to adopt AI without adequate training. Anthropic is also updating Claude’s usage policies to address abusive interactions, deceptive campaigns and dangerous autonomous systems. AI is moving into increasingly consequential roles. Utah is launching a limited pilot allowing AI to assess patients with mild-to-moderate acne and recommend prescriptions, with medical oversight gradually reduced as the program progresses. Military leaders predict AI could coordinate networks of battlefield sensors, drones and weapons by 2030, raising difficult questions about human control over autonomous warfare. Meanwhile, Google introduces a universal Gemini enterprise agent that can work across business applications, query data across cloud platforms and select between Google and Anthropic models. Microsoft is also upgrading Copilot’s Audio Overview, turning emails, meetings and notes into audio summaries with improved navigation and data protection. Plus: a man who used AI-generated music and more than 1,000 streaming bots to defraud music services of over $8 million** is sentenced to 18 months in prison. A smart coffee maker reportedly uploads more than 1 TB of data in 10 days, highlighting the security risks of connected household devices. Microsoft reduces shared OneDrive storage for Microsoft 365 Family and Premium subscribers from 6 TB to 2 TB, prompting customer backlash, while a developer uses AI to build seven free, open-source alternatives to Adobe applications. OpenAI introduces faster, more expensive processing for GPT-6.1 Sol and a new Decisions API designed for rapid, structured decisions. In crypto, Ethereum founder Vitalik Buterin warns that AI-driven mathematical breakthroughs could threaten existing cryptographic methods sooner than expected. On the business and economic front, Anthropic could pursue an IPO valuing the company at approximately $2 trillion, potentially raising as much as $100 billion. Google signs a 20-year agreement supporting 890 MW of additional nuclear capacity, reinforcing the growing connection between AI data centres and electricity demand. Samsung forecasts quarterly operating profit of approximately $80.2 billion**, driven by surging memory demand, while Paramount Skydance completes its $110 billion acquisition of Warner Bros. Discovery. In Canada, September unemployment rises to **6.5% after 68,000 jobs are lost, Toronto home sales fall 9% year over year, and business leaders call for tax reforms to encourage domestic investment. Finally, Apple may unveil its first touchscreen MacBook and other hardware updates later this month.