A house newsletter · daily · researched and written by the Kurzia AI
Sunday 15 November 2026 · Issue 14 · 3 min read
Lovable apps expose passports and chats; an agent co-sells six-figure deals; Blackwell fingerprints in Chinese weights
The story that matters most
Crissy Field GmbH · yesterday
Lovable apps ship customer data to anyone who asks
A thin weekend calendar, but one report deserves the morning. Security firm Crissy Field, working with Handelsblatt, analysed 159,041 Lovable-built projects found via Common Crawl between December 2025 and September 2026; 34,545 used Supabase, and 24,664 of those sites served database records on first load, no login required. Pulling further records with the same public API key the sites already hand out, the researchers found email-bearing rows in 10,560 tables across 7,403 sites — about 30% of those closely examined, which the firm stresses is not a confirmed leak rate. The worst finds: a German forum exposing over 80,000 profiles and more than 200,000 chat messages, and a provider leaking over 1,000 records with passport numbers.
Lovable — last valued at $13.3bn on $400M ARR, reported by Trending Topics — markets that no technical knowledge is needed, while its terms put data handling on customers, and because its database setups repeat a pattern, one working attack scales across thousands of apps. Any portfolio company shipping an AI-built app with real customer data needs an expert review of access permissions before launch, and security tooling for vibe-coded apps is now a category worth a look. Crissy Field's own warning is the diligence standard: automated scans are no substitute for a professional review of permissions, interfaces and account separation before an application touches real data.
“An easy-to-use development platform does not remove the need for an expert review of access permissions before an application handles confidential data.”
For your portfolio
SaaStr · yesterday
10K
Takeaway: from the 20s to the 70s
SaaStr's Jason Lemkin reports that 10K, the agent his company bills as its VP of Marketing and Revenue, co-closed two six-figure sponsorship deals in one November week — one with Base44, Wix's AI app builder — while still unable to run a deal alone: no live calls, no negotiation, nothing signed. 10K built the competitive deck during the inbound session, priced custom booth work from 13 years of invoices the same afternoon, and prepped procurement and vendor paperwork; a human took every call and approved every Salesforce write.
This is the clearest public account of the co-sell split: agents now cover research, custom pricing, procurement and CRM hygiene, leaving people the live conversation. SaaStr's own numbers: renewals for its $100K+ sponsors went from the 20s to the 70s this year, roughly tripled in the 11 months since 10K started — a correlation the company does not claim to fully attribute to the agent. Ask a founder: let an agent run everything but the calls on the next five deals needing custom pricing, and count the days saved. Reps still spending a week digging through old quotes are behind.
Unemployment Labs · 3 days ago
Chinese frontier weights carry Blackwell fingerprints
A forensic report by Will DePue's Unemployment Labs, published November 12 and still driving coverage through the weekend, reads the published weights of six Chinese models released since April — from DeepSeek, Moonshot AI, MiniMax and Ant Group — and finds their initialisation patterns match Nvidia's export-controlled Blackwell chips, while earlier models from the same labs match the older Hopper generation. The method, validated against 24 models with disclosed hardware, shows what the weights were started on; it does not establish where the chips sat or that export law was broken. The report puts the five final training runs at more than half a billion dollars of Blackwell hardware.
For portfolios building on Chinese open-weight models because they are cheap or free: that cost advantage partly rests on hardware US export rules meant to keep out of China, so a licence change, crackdown or takedown could strand products built on those weights. Ask founders running on them: if DeepSeek-V4 or Kimi K3 weights were pulled tomorrow, which products break, and what is the named fallback?