A security research firm sampled 159,041 vibe-coded Lovable apps and found passport numbers, payment data and private chats readable without a login - while the Wall Street Journal reconstructs how Nvidia’s financing machine met its “avalanche.” Plus the quiet ones nobody talks about.
← All issues
The Diff  No. 043
Sunday, 15 November 2026  ·  AI dev news, only what is verified
Google Research artwork for Antigravity’s /teamwork multi-agent feature
A security research firm sampled 159,041 vibe-coded Lovable apps and found passport numbers, payment data and private chats readable without a login - while the Wall Street Journal reconstructs how Nvidia’s financing machine met its “avalanche.”

Crissy Field, a German security firm working with Handelsblatt, pulled database records out of thousands of Lovable-built sites using nothing more than the public API key each site ships to every visitor - and says 30% of the sites it checked returned records containing email addresses.

And the Journal’s account of Jensen Huang’s final months describes a financing program that backstopped customers who could not afford Nvidia’s chips, a CFO who asked how to value the guarantees, and a CEO who called her “a deer in the headlights” months before she took his job.

Two flavors of the same lesson: the easy path ships risk by default. Vibe-coded apps leak when nobody reviews access rules, and easy credit leaks into prices when nobody asks who is guaranteeing the guarantee.

IN THIS ISSUE
01
🔓 Vibe-coded Lovable apps are exposing passports, payment data and 200,000 private chats
Security firm Crissy Field sampled 159,041 Lovable-built projects and pulled records containing email addresses from 30% of the sites it examined closely - using only the public Supabase key each site already hands to every visitor.
🏔️ WSJ: Huang saw an “avalanche” of chips coming - and berated the CFO who now has his job
02
🏔️ WSJ: Huang saw an “avalanche” of chips coming - and berated the CFO who now has his job
The Wall Street Journal reconstructs Nvidia’s customer-financing spree: credit guarantees for buyers who could not afford the chips, a $250 billion OpenAI backstop halved by Colette Kress, and a soup-vendor fable in Berlin. Reporting, not filings.
🔥 VIRAL
The loud ones on Hacker News. Numbers are a morning snapshot.
01  crissyfield.de research, with Handelsblatt · HN 703 points, 290 comments at 06:26 IST

🔓 Vibe-coded Lovable apps are exposing passports, payment data and 200,000 private chats

Security firm Crissy Field sampled 159,041 Lovable-built projects and pulled records containing email addresses from 30% of the sites it examined closely - using only the public Supabase key each site already hands to every visitor.
SCENE

A German lifestyle forum had more than 80,000 user profiles - full names, email addresses, dates of birth, gender - plus over 200,000 private chat messages readable by anyone, no login needed. Elsewhere the researchers found ticket and food orders with payment information, AI-generated summaries of a property manager’s internal emails, and, in their words, a particularly serious case: over 1,000 records with identity-document details including passport numbers, expiration dates and nationalities.

THE THING

The verified figures, all from Crissy Field’s own report. They identified 159,041 Lovable projects in Common Crawl data from December 2025 to September 2026; 34,545 used Supabase. Of those, 24,664 sent database records on first page load - not automatically a problem, since some data is meant to be public. But with further read-only requests using the same public API key, they found records containing email addresses in 10,560 tables across 7,403 of those 24,664 sites - 30%. The authors themselves warn this is not a confirmed leak rate: some records may be public contact data, and their Common Crawl sample is not representative. They downloaded only parts of datasets, publish no access details, and could not establish whether any data was misused. Lovable’s terms place responsibility for applications and data on customers; on November 10, after Handelsblatt’s first story, Lovable published its own security guide, “Find it, fix it, ship it.”

✂️ STEAL IT

Steal the five-minute audit. Supabase’s anon key is public by design - the only thing protecting a table is its row-level security policy. For any app you have shipped with Lovable, Bolt or a Supabase backend: open an incognito window, watch the network tab on first load, and see what your own site hands a stranger. Then confirm every table has an RLS policy that assumes the key is public - because it is. If the app ever touches real customer data, get a human security review before launch, not after a researcher emails you.

Try today  VERDICT

The audit habit costs five minutes per app and the report’s methodology section tells you exactly what researchers - and attackers - look for. This fails you if you read “no technical knowledge needed” as “security included”: the platform builds the app, but the access rules are yours, and Lovable’s own terms say so.

Crissy Field: the research report  ·  Lovable’s own security guide (Nov 10)  ·  HN discussion

Google Research artwork for Antigravity’s /teamwork multi-agent feature
02  WSJ (Gardizy & Whelan, Nov 12) · TechSpot follow-up at HN 317 points, 173 comments at 06:26 IST

🏔️ WSJ: Huang saw an “avalanche” of chips coming - and berated the CFO who now has his job

The Wall Street Journal reconstructs Nvidia’s customer-financing spree: credit guarantees for buyers who could not afford the chips, a $250 billion OpenAI backstop halved by Colette Kress, and a soup-vendor fable in Berlin. Reporting, not filings.
SCENE

This summer, Jensen Huang summoned his finance chiefs to diagnose a ballooning problem: Nvidia was in talks to guarantee billions for companies that could not afford its chips. Mid-presentation, he interrupted Colette Kress - named interim CEO on Thursday - calling her “a deer in the headlights,” according to a person familiar with the comments. The committed chip and memory supply coming at them, Huang said, was an “avalanche.”

THE THING

What the Journal reports, attributed to people familiar: the financing push was led by Huang’s son-in-law Nico Caprez, promoted to VP in January, with VP Raj Mirpuri, promising credit support within weeks in exchange for a revenue cut. For OpenAI’s Ohio campus - roughly $300 billion of Nvidia chips - Nvidia weighed a $250 billion backstop; the day after WSJ first reported that figure on July 26, Nvidia shares fell 5%, erasing $250 billion in market value. Kress reworked the deal and halved Nvidia’s commitment; the downsized deal signed August 17. Kress publicly defended the program in August: “We recognize the scale of this support, and we know some will call this circular financing. We see it differently. The equity returns on our invested capital will be excellent.” By October, Nvidia “started to turn off the tap,” and in Berlin Huang told chip-buyer CEOs a fable about a soup vendor with too many customers and not enough soup. Separately verified at the primary source: director Mark Stevens sold about $946 million in Nvidia stock in August and September; the SEC Form 4 for the largest batch (1,356,000 shares at $219.72 on Sept 18) is on EDGAR. There is no evidence connecting the sales to the resignation; TechSpot’s headline claim that Nvidia “deliberately kept GPU prices high” is that outlet’s inference, not the Journal’s wording.

✂️ STEAL IT

Steal the supply-chain read: the AI capacity you rent is partly financed by the company selling the chips, and the financier is now tapping the brakes. If your roadmap assumes falling GPU cloud prices, stress-test it - between wafer “avalanches,” consumer VRAM diverted to data centers, and paused credit deals, cheap abundant compute is a bet, not a baseline.

Watch  VERDICT

Nothing to deploy or patch. The first on-the-record checkpoint is Nvidia’s Q3 earnings call on Tuesday, November 17 - listen for whether Kress takes questions on the financing book and the investigations. Treat every vivid detail here as WSJ reporting attributed to unnamed sources until a filing or an official statement confirms it.

TechSpot’s report on the WSJ profile  ·  SEC EDGAR: the Stevens Form 4 (Sept 18 batch)  ·  AP: why Kress is stepping in

📊 THE ATTENTION GAP
HN points at 06:26 IST. Attention, not quality.
SpaceX catches Starship, Flight 15
887 HN
Lovable apps expose data
703 HN
Nvidia circular-financing report
317 HN
ALSO ON THE RADAR

Google Research says its Antigravity /teamwork agents helped prove a conjecture open since 2013. In a Nov 13 post, Google Fellow Vahab Mirrokni says researchers using Gemini 4 models with /teamwork multi-agent workflows proved information theory’s Courtade-Kumar conjecture in full generality and formalized the proof in Lean 4 - producing what the post calls 50 million lines of code - plus results on open questions in adaptive data analysis (open since 2014), quantum query complexity (2014), Boolean matrix multiplication (STOC 2024) and streaming cut sparsification (FOCS 2022). The patterns, including the /boost solve-and-verify loop, are “available in Antigravity today,” with fuller external rollout “in the coming weeks.” All results are Google’s own claims; independent verification will take the math community some time. Google Research: /teamwork results →

Decider v1.3 lands on Artificial Analysis’ Decision Index at 70 - one day after release. The independent score for Thursday’s lead story: Decider v1.3 places #6 of 22 decision models, above the median of 66, at $68 per million tasks - but slow, at 289 ms per task versus a 209 ms median. Decider Flash v1.3 scores 64. This is the first third-party benchmark of Perplexity’s decision model; the release-day numbers in No. 042 were Perplexity’s own. Artificial Analysis: Decider v1.3 →

Serving GLM-5.3 takes two DGX Stations, and Nvidia shows the wiring. Nvidia’s technical blog walks through serving GLM-5.3 - a 753-billion-parameter MoE, 40 billion active, needing about 464 GB in NVFP4 - across two DGX Stations linked by ConnectX-8 RDMA, behind one OpenAI-compatible vLLM endpoint with TP=2. Its AIPerf benchmark peaks at 607.8 output tokens per second at concurrency 16, falling to 385 at concurrency 32. Vendor benchmarks, on Nvidia’s own hardware. Nvidia Technical Blog: GLM-5.3 on two stations →

PyPI’s redesigned project pages are live, with security signals up front. The Python Package Index finished its UI rollout: project pages now surface security information - including vulnerability advisories and project-health signals - directly on the page instead of buried in metadata. Live since the Nov 6 announcement; the HN discussion (188 points) is from Saturday. PyPI blog: the redesign →

🧪 ONE THING TO TRY TONIGHT
Five minutes: check what your app hands a stranger.
Shipped anything with Lovable or Supabase? Open the live site in an incognito window, open the network inspector, and reload. Any Supabase request you see is one anyone on the internet can replay - the anon key is right there in the page. Copy one request, strip your login, and run it: if rows come back that should need an account, that table is missing a row-level security policy. Fix the policy before you fix lunch. While you are at it, Lovable’s own Nov 10 security guide walks the same audit from inside the platform.
What this issue could not see. The Lovable figures - 159,041 projects, 34,545 on Supabase, 24,664 sending records on load, 7,403 sites (30%) with records containing email addresses - are Crissy Field’s, from its own report; the authors state the 30% figure is not a confirmed leak rate and their sample is not representative, they did not download full datasets, and they found no evidence of misuse. The Nvidia account is Wall Street Journal reporting attributed to unnamed “people familiar”; the only board-confirmed facts are Huang’s resignation and Kress’s appointment (No. 041), and the Stevens stock sales are verified only as to the SEC-filed Form 4 transactions - no evidence connects them to the resignation, and TechSpot’s “deliberately kept GPU prices high” framing is that outlet’s inference. The /teamwork research results are Google’s claims, not independently verified; the Decider v1.3 Decision Index score is Artificial Analysis’ independent measurement; GLM-5.3 throughput numbers are Nvidia’s own benchmarks. Follow-ups on Friday’s Anthropic outage day: no post-mortem or cause has been published, and a new, minor incident - Sign in with Apple failing for new iOS accounts - has been open since Nov 14, 07:11 UTC. The FAA ground stop’s root cause is still officially unknown: contractor L3Harris confirmed it is converting airspace communications from copper to fiber (70% complete) and is supporting the FAA investigation; the FAA says the cause is unknown, while the Transportation Secretary suggested L3Harris’s work “may have” contributed - an unresolved attribution split, and still off this newsletter’s beat. VIDEO OF THE DAY is omitted for the tenth day running: no fresh, relevant and verifiable video on today’s stories could be confirmed. Reddit remains absent by policy; X blocks anonymous access and was not consulted. Deduped against every published issue, No. 001-042, via the kit shipped-index and the live home archive: the Lovable exposure, the WSJ Nvidia account, /teamwork results, GLM-5.3 serving and PyPI’s redesign are new; Decider v1.3’s release was No. 042 and appears only as the independent-score follow-up.
Images: credited in each report; source screenshots and announcement artwork, not our tests.