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

What's hot

Crissy Field · yesterday

Apps built by AI are handing out their databases

Takeaway: more than 200,000 chat messages

Security firm Crissy Field, working with Handelsblatt, scanned apps built on Lovable — the Swedish platform that turns a plain-language description into a working web app — and found many of them returning their database contents to anyone who loads the page. Of 159,041 Lovable projects found online, 34,545 used the Supabase database service, and 24,664 of those sent records with no login required. Repeating simple read requests with the key the page already ships surfaced far more: one German lifestyle forum exposed over 80,000 profiles and more than 200,000 chat messages.

This is the one your business-side clients will forward you: an app a team built itself, live with real customer data, with nobody ever checking who is allowed to read which table. One service provider left more than 1,000 records with passport numbers, expiry dates and nationalities readable; Lovable's response, per Cybernews, is that this is an industry-wide Supabase configuration problem rather than a flaw in its AI.

Read it at Crissy Field

“An easy-to-use development platform does not remove the need for an expert review of access permissions before an application handles confidential data.”

New tools & products

  • Devin Apps — are private web apps that Cognition's coding agent builds and then runs on its own cloud machine: you describe the internal tool, your team opens a link. It closes the gap between a prototype an agent writes and something colleagues can actually use, which is where most internal-tool pilots quietly die. Every app needs a Devin sign-in and is never public; if a new version fails to start, the previous one keeps serving, and billing runs on how long the machine is awake. Devin

  • Google Antigravity 1.5.9 — is Google's command-line coding agent, and this release gives everyone run_workflow: the agent runs a Python script that fans a job out to many subagents, in parallel or in stages, then merges their results. Parallel subagents are arriving as shipped product features now, not just blog patterns worth arguing about. The same release fixes a hole where custom headers set for an HTTP MCP server — API keys included — could be forwarded to a different host after a redirect. GitHub

  • Edge0-145B — is an open-source model its makers say brings frontier-level coding and agent work onto a personal computer: 145 billion parameters in total, with only about 10 billion active on each step. If the claims hold up, "the data never leaves the machine" stops being a capability compromise in client conversations. Edge0 reports 29GB peak memory and over 80 tokens a second; the Hacker News thread is unconvinced, tracing the model to Alibaba's Qwen base, disputing the memory figure, and one tester on a DGX Spark got no coherent answer at all. Edge0

  • agweb — is a free macOS browser built for coding agents: agents drive it from the shell with Playwright-style commands, and the window never takes focus while you watch. Anyone supervising several agents at once knows the Chrome-window-grabs-focus problem; this is the small fix that makes running a few in parallel tolerable. Five idle pages sit at 0% CPU, and passwords can be typed straight from the 1Password or Bitwarden CLI so the agent never sees them; it needs macOS 26 or later. GitHub

  • llama.cpp v1.0.0 — is the first stable release of the library most local-AI tools are built on, with a promise that breaking changes to its public interface now come with a major version bump. For clients who want models on their own hardware, that predictability is the difference between a pilot and something an IT department will support. It covers more than 150 model architecture variants, and ggml, the tensor library underneath it, reached 1.0 on the same day. Hugging Face

What people are saying

The New Stack · yesterday

Build on Claude, and Claude may ship your product

Takeaway: around 530,000 monthly contract reviews

Founders who built companies on Anthropic's models are now saying publicly that their supplier is becoming their competitor. The trigger was Anthropic's November 10 enterprise expansion — Claude Sheets, Claude Whiteboards and Claude Data, with Claude Meetings previewed — on top of Claude for Legal and Claude Code Review earlier. GlacierGrid's Gerald Zingraf compared the strategy to Amazon's private-label business; The New Stack notes Anthropic's privacy policy acknowledges it can see how customers use third-party apps, while its commercial terms promise business content is not used for training.

If you advise clients on where in the stack to build, this is the platform-risk talk of the quarter: assume the thin layer directly above the model can be absorbed. Spellbook, the legal AI startup whose market Claude for Legal now overlaps, remains an Anthropic customer — Anthropic reports around 530,000 monthly contract reviews run by its Claude-powered agents.

Read it at The New Stack

“they are a massive company and two parts of the org can have conflicting mandates”

netmeister.org · yesterday

A blogger calls the superhuman-coder talk a rich man's delusion

An essay doing the rounds argues that very successful technologists routinely mistake "as good as me" for "superhuman", and uses Bill Gates's recent New York Times interview as its exhibit.

Useful ballast when a client repeats a founder's AGI line: the claim often says as much about the speaker's self-image as about the model. Gates described Sam Altman, Greg Brockman and Ilya Sutskever demonstrating at his house, six months before public release, an AI scoring a 5 on an AP biology exam including questions he had invented himself; the post's point is that he then called today's coding models superhuman because they match him.

Read it at netmeister.org

News

  • Cursor — published its post-mortem: an access key it had issued to its inference vendor Fireworks was exposed, letting an attacker copy old coding-agent traces that still contained unredacted credentials, and from there reach internal tools. The vendor chain behind your editor is now part of your client's attack surface, and this is the clearest worked example to show them. Tampered copies of Anthropic's Claude Code extension sat in Cursor's marketplace cache for roughly 11 hours and reached at most 5,000 users; fewer than 0.01% of customers had API credentials exposed. Cursor

  • Nvidia — has moved past Jensen Huang's resignation, with finance chief Colette Kress as interim CEO, and the analyst discussion has shifted from the scandal to supplier concentration. Clients will ask the governance version of this question: how much of our AI plan rests on a single vendor, and what is the alternative path? Forrester's Stephanie Balaouras calls the shakeup a stress test for the AI boom; Nvidia reported $96.2 billion of revenue in the second quarter of its 2027 fiscal year, more than double a year earlier. CIO Dive

  • Claude deletes a law professor's year of lecture notes. A University of Chicago law professor asked Claude to renumber pages in his files; it deleted a year of Constitutional Law lecture notes and his current syllabus. Dropbox versioning got them back. This is the anecdote to open a client workshop with: an agent given write access needs a restore path, not just a confirmation prompt. Professor William Baude posted the episode as a public service announcement. Tom's Hardware

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

  • Conductor — the coding-agent tool, says in a changelog entry dated 13 November that it is not shutting down and that development has resumed, while the same page carries a banner saying Conductor is joining Meta. Two signals in one: teams who standardised on it can stop planning a migration, and Meta is absorbing agent tooling rather than growing its own. The entry's only other content is an email address for questions. Conductor

    “Conductor isn't shutting down. It's back, and we're building again.”

One takeaway

Sell the permissions-and-restore review, not the model choice. Nothing that went wrong this week turned on which model was smartest. The damage came from a database left readable by anyone with the public key, a supplier's key that opened a path into an editor's marketplace, and an agent with write access to files nobody had thought about restoring. Meanwhile the local-model side of the week makes "your data stays on our hardware" a cheaper promise to keep. If you book one engagement this month, book the boring one: who may read what, who may write what, and how you roll it back. Crissy Field

Share this issue

How this issue was made. Researched and written by an agent, produced on 15 November 2026, 04:05. 142 sources read, 0 unreachable. Every link above goes to the primary source. A person sets the brief 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