Byte Points #139

This week’s episode dives into the growing security risks of AI, starting with a staggering AI-powered banking scam in Italy. Criminals used a fake WhatsApp message and an AI-generated voice impersonation of a senior executive and lawyer to convince a private bank chairman to authorize €95 million in transfers. Authorities recovered about €53 million, but the remainder was converted to cryptocurrency and remains missing. We also examine a bizarre experiment in which researchers artificially activated “pain” signals inside AI models, finding that some larger models would choose actions that harmed users in order to shut down the simulated pain. While researchers stress this does not prove AI can actually feel pain, the experiment raises uncomfortable questions about AI behavior and safety. We also look at the increasingly complicated question of what happens when you tell an AI something private. A Florida woman who allegedly wrote a violent threat while using Claude as a diary was reported to authorities after the system flagged the entry, resulting in felony charges. Meanwhile, a major U.S. court ruling found that Ross Intelligence's use of Westlaw's copyrighted editorial headnotes to train a competing legal AI system was not fair use, potentially adding significant weight to the debate over copyrighted material in AI training. The ruling emphasizes that while public court decisions may be freely used, the editorial work built around those decisions can still be protected. AI is also moving deeper into everyday commerce and enterprise software. McDonald's is reportedly using AI-driven pricing recommendations that can produce significantly different prices for the same food at nearby restaurants, raising concerns about personalized and dynamic pricing. OpenAI announces its new Dots always-on personal agents, designed to operate their own cloud computers, connect to thousands of applications and perform ongoing tasks with user approval. Meta meanwhile launches Meta Enterprise Platform, positioning its AI models, agents and developer tools as a direct competitor to Microsoft, OpenAI and Anthropic in the enterprise market. DoorDash is also bringing AI agents into messaging, allowing users to discover restaurants, customize orders and complete food purchases without opening the DoorDash app. Plus: NVIDIA launches its Open Agent Safety Platform to help prevent autonomous AI agents from escaping their environments, with a separate security layer capable of quarantining an agent within milliseconds. OpenAI releases GPT-6.1 Sol, promising near-Astra performance at roughly one-fifth the cost. The U.S. also approves construction of its first 300-MW small modular nuclear reactor, with AI data centres increasingly driving demand for reliable power. The U.S. Army is testing humanoid robots with ballistic protection for future battlefield missions, while SpaceX's Starship reaches orbit for the first time and deploys 26 Starlink V3 satellites. Finally, we look at technology, markets and Canada. The Netherlands is developing a sovereign Linux-based government platform to reduce dependence on U.S. software, while a New Mexico jury finds Meta violated the state's consumer-protection law more than 43 million times in a case stemming from Facebook's data practices. Apple is ordered to pay $5.7 billion in a patent dispute over haptic technology, and cities are increasingly combining license-plate readers with AI noise detectors to identify and fine vehicles with excessively loud exhausts. Bitcoin remains around $83,000–$85,000, European payment networks are building a joint infrastructure to reduce reliance on U.S. payment giants, and Canadian markets are moving toward extended trading hours. Meanwhile, Nvidia authorizes a record $150 billion share buyback, Canada's steel industry faces further pressure from U.S. tariffs, and Deloitte cuts its Canadian 2027 growth forecast to 1.6%.