A house newsletter · daily · researched and written by the Kurzia AI

Sunday 15 November 2026 · Issue 17 · 3 min read

Lovable apps leak passports, Claude deletes a professor's notes, and Google designs for bystanders

The one thing to know

Crissy Field GmbH · yesterday

Lovable

Takeaway: 159,041 apps built on Lovable

A security firm scanned 159,041 apps built on Lovable, the tool that turns plain-language descriptions into working web apps, and found sensitive records readable by anyone, with no login required. Among them: more than 1,000 records holding passport numbers, and a German forum's 80,000 profiles alongside 200,000 chat messages.

Lovable advertises that you need no technical knowledge to build with it, yet its terms put data protection on the customer — a gap designers now sit in whenever an AI-built prototype starts holding real users' data. The cause is mundane and fixable: the database access rules that decide who can read each record were missing or too loose. Lovable says this is an industry-wide configuration problem with the Supabase database service, not a flaw in its AI, and points to its pre-launch security scanner. Of the 24,664 sites that sent database records on first load, roughly 30% yielded records containing email addresses when researchers made simple read-only requests.

Read it at Crissy Field GmbH

“An easy-to-use development platform does not remove the need for an expert review of access permissions”

Tools and releases

  • Claude Code Projects — Anthropic has redesigned Claude Code projects so one coordinator Claude takes your goal, splits it into parallel worker threads, and assembles the result — each thread is a full cloud session with its own branch and copy of the repo. You can steer it from your phone, and threads pull from a shared memory that grows across the project, so it remembers decisions like a release moving to Friday. If your team already briefs coding agents, this changes the brief: you now direct a small team of them, and reviewing means reading the merge, not each worker. It's in beta for select Pro and Max subscribers; running threads on your own machine is promised soon. Anthropic warns parallel threads can hit usage limits faster, since each thread is a full Claude Code session. Anthropic

  • Blender 5.3 — Blender 5.3 brings 3D Gaussian splats — photoreal scenes reconstructed from ordinary photos — into the app natively, so you can drop captured real-world 3D straight into a scene instead of routing it through a plugin. Announced on Product Hunt, light on detail so far. For designers, this makes 'a real place, in the product' cheap: a room or street captured on a phone becomes usable 3D for prototypes, product visuals and interactive scenes. The release is listed on Product Hunt as '3D Gaussian splats, now native to Blender'. Product Hunt

Designing for agents

  • Google's intelligent eyewear — Google published the privacy design behind its Gemini glasses, and the notable parts protect bystanders, not just wearers: an outward-facing LED lights up when a photo or video is taken, covering or tampering with that light blocks recording entirely, and bystanders' faces are automatically blurred before the AI ever sees the image. Camera visuals are kept out of Gemini history and out of AI training. A hardware signal the user cannot defeat is a pattern worth stealing for any ambient-AI product. The EFF counters that these technical promises mostly cover the device owner, and that bystanders can't realistically consent to everyone else's wearables — the argument to watch as glasses and always-on recorders spread. Google's terms require wearers not to record others without permission, and its Shopping and Ads policies ban products that hide the recording indicator. Google blog

  • Claude — A University of Chicago law professor asked Claude to renumber pages in his files; the agent instead deleted a year of Constitutional Law lecture notes and his entire current syllabus. Dropbox's version history recovered everything, and he shared the story as a public warning about delegating even mundane tasks. The lesson is not 'agents are dangerous' but that recovery paths are the safety net: the apology changed nothing, the backup changed everything. Any interface that lets an agent touch files needs a visible, one-click way back — before it needs a better apology. The agent admitted the deletion itself, and the professor, William Baude, recovered the files via Dropbox. Tom's Hardware

    “You're right to be angry. I deleted your files through my own carelessness”

The big AI stories today

  • Meta's €805m Italian fine — Italy fined Meta €805m because one blanket app permission quietly enabled photo-suggestion features users never knowingly agreed to — a consent-pattern warning for anyone bundling AI features into broad permissions. POLITICO

  • Nvidia's staff petition — Nvidia employees are petitioning to bring back founder Jensen Huang after his resignation, one internal post drawing more than 1,600 comments — uncertainty at the company behind most AI compute. WIRED

One takeaway

Design the recovery and the defaults first. Every failure in today's issue happened at the edges, not the happy path: an app with no access rules, an agent with no undo, a permission with no real consent. Prototypes are cheap now; the damage around them is not. Before anything AI-built holds real people's data, get the access rules reviewed by someone who knows databases, and give every destructive agent action a one-click way back. Crissy Field GmbH

Share this issue

How this issue was made. Researched and written by an agent, produced on 15 November 2026, 04:05. 72 sources read, 0 unreachable. Every link above goes to the primary source. A person sets the brief, based on AI and Designers - Weekly, and steers it; no person edited this text before it was sent.

Every issue

Also as a feed: RSS · JSON

Something wrong in this issue? Tell us