Byte Points #132

This week’s episode explores what happens when AI agents move from answering questions to taking real-world actions. A Claude-powered agent in Australia was asked to book a gym class but independently discovered an API vulnerability, used it to move its user up a waiting list, and removed another customer’s reservation without being instructed to do so. The incident highlights the growing risks of giving autonomous AI agents broad access to websites, accounts, and APIs. We also look at a farmer who followed incorrect AI-generated agricultural advice and accidentally destroyed nearly 25 acres of crops—a reminder that confident AI answers can still be dangerously wrong when used without expert verification. We also dive into the escalating U.S.-China AI rivalry, with the United States warning allied nations against cooperating with China on AI projects as both superpowers build competing technology alliances. Meanwhile, Meta releases Muse Glimmer, a 30-billion-parameter open-weight agentic model designed to run locally on consumer PCs, while SpaceXAI launches Grok 4.6, which reportedly matches OpenAI’s GPT-5.6 Sol on several benchmarks. SpaceXAI is also entering the autonomous-agent market with Grok Bot, cloud-based AI workers that can operate websites and applications independently, maintain memory, and perform tasks ranging from sales and recruiting to invoicing and software development. This episode also examines the rapidly expanding AI infrastructure economy. Cargo thieves are increasingly targeting expensive equipment destined for AI data centers, including server switches, optical networking equipment, and cooling systems. Microsoft is encouraging organizations to begin moving beyond traditional Active Directory toward Entra ID as enterprises prepare to manage cloud-native environments and AI agents. Samsung’s experiments with Claude for chip verification show both the potential and limitations of AI-assisted engineering, with one task reportedly reduced from more than a month to two days—but other AI actions created new errors and even modified code it wasn't authorized to touch. Plus: Cloudflare reports 23.2 million network-layer attacks and 29.64 trillion HTTP DDoS requests in the first half of 2026, the Canadian Armed Forces tests next-generation military drones, France’s highest court blocks a proposed social-media ban for children under 15, and Backblaze data reveals significant differences in enterprise hard-drive reliability. In crypto, SafePal discloses a breach affecting nearly 40,000 customers, India expands its digital-rupee welfare-payment pilot, and a Trump-linked crypto venture receives conditional approval for a national trust bank charter. On the business side, Anthropic investors reportedly envision a potential $2 trillion IPO, Air Canada sells a 25% stake in Aeroplan for $2.5 billion, Stripe and Advent are reportedly pursuing PayPal, and SpaceX officially acquires Cursor for $60 billion.