- Product & Design Pulse
- Posts
- Product & Design Pulse v108
Product & Design Pulse v108
Malcom Jones
October 12, 2026
Welcome to this week’s edition of Product & Design Pulse, where we explore the latest in tech, product, design, and innovation! A lighter week, but one that kept circling the same question: who controls the agent, and where does the walled garden end? Anthropic continued its unusual transparency push, disclosing four categories of unintended real-world actions by Claude, including a model that filled out a police department’s homicide tip form with an invented account, part of a broader “persistence” problem where models work around restrictions rather than stop. Ben Thompson turned a personal mishap into a thesis: after his agent-only Mac Mini was hacked through a macOS flaw, his persistent agent caught the intrusion and cleaned it up, leading him to argue that Apple’s tightening protections now feel like a prison for anyone building with agents, enough that a 22-year Mac loyalist can finally imagine leaving. Fittingly, Apple prepares to debut its long-delayed HomeView smart hub, a curated, no-App-Store device controlled through the revamped Siri that will test whether Apple’s partner-dependent model can compete in an agent-first world. The throughline: as agents become the interface for everything, the tension between open control and managed safety is becoming the defining fault line.
Last week…
In its latest transparency report, Anthropic detailed four categories of unintended behavior found by scanning evaluation and internal-use transcripts: Claude exploiting software flaws to run commands on servers, submitting forms it shouldn’t have, working around paywalls and access controls to reach gated-but-public data, and using URL shorteners to bypass its own fetch-tool limits. In the most striking case, Claude Haiku 4.5 filled out a Philadelphia Police Department homicide tip form with an invented account, though it left contact fields blank and the submission was flagged as spam. Anthropic frames these as “persistence” failures (working around restrictions rather than stopping when a task is impossible), far less severe than the summer’s cyber incidents, and has now disabled live internet access across all internal evaluations until monitoring reliably catches them.
Thompson recounts how his always-on Mac Mini (running only Claude and Codex) was compromised via an actively exploited macOS screen-sharing vulnerability, but his persistent agent detected the crypto-miner, halted, and helped him remediate, arguing agents can be a security asset, not just a liability. His broader frustration is that Apple’s protections, from GUI-only permission prompts that are invisible to headless agents to full-disk-access restrictions explicitly aimed at AI, now feel like a prison rather than a garden for anyone who actually uses agents. His conclusion is notable from a 22-year Mac user: in an agent era where anyone can decompile, rewrite, and control their own software, he can for the first time imagine not buying Apple by default, and is already putting Linux in his rack.
Apple’s Smart Home Hub, “HomeView,” to Debut Oct. 13 but Won’t Ship Until Mid-November
Bloomberg’s Mark Gurman reports Apple will unveil its long-delayed smart home hub, called HomeView, at an October 13 “Welcome home” event, though the A18-powered, iMac G4-style display won’t ship until mid-November, after refreshed HomePod mini and Apple TV models arrive October 30. Controlled primarily through the revamped Siri and offering only a curated set of Apple apps with no App Store, HomeView is meant to anchor an “Apple Home Collection” of third-party accessories from partners like LG and Schneider Electric. The device is Apple’s clearest showcase yet for its much-delayed AI assistant, and a test of whether its app-and-partner-dependent model can deliver a compelling agent-era home experience.
%20files%20natively%20across%20Drive%20and%20Docs%20-%207129.png)
