Leaky AI-built apps, a 16-token memory for LLMs, and a shopping show with no humans
Today's stories are about what happens when AI does the building. A study found thousands of apps made with the AI builder Lovable handing out database records to anyone who asked. Stanford researchers released a small open module that squeezes a pile of documents into 16 tokens, and a Korean TV retailer is airing a sales show with no host and no camera.
Key takeaways
- Researchers found email addresses readable without a login on 7,403 Lovable-built sites, so anyone who has shipped an AI-built app on Supabase should check its access rules before it holds real customer data.
- Stanford's Memorilla gives a frozen Qwen3-8B model a 16-token memory that beat full-context prompting on most of the paper's tests, but it lost clearly on PubMedQA and needs an 80 GB GPU, so test it on your own data first.
- Hyundai Home Shopping says a fully AI-made sales broadcast cut suppliers' production costs by more than 80% in a trial, a company figure with no detail behind it, but a clear sign that small sellers will want cheap product video.
Study finds thousands of Lovable-built apps sending database records to anyone, including passport details
Hamburg security firm Crissy Field, working with the German newspaper Handelsblatt, examined 159,041 projects built with the AI app builder Lovable and found email addresses readable without a login on 7,403 sites. Exposed records included more than 200,000 private chat messages and more than 1,000 entries with passport details.
Crissy Field, a security and software consultancy in Hamburg, has published a study of apps built with Lovable, the Swedish tool that turns a text prompt into a working web app. Working with Handelsblatt, which first reported the findings on 2 November, the firm identified 159,041 Lovable projects in the Common Crawl web archive, covering December 2025 to September 2026. Its English report, which uses updated figures, drew wider coverage this weekend. The core finding is that many of these apps hand database records to any visitor, with no login.
The numbers narrow in steps. Of the 159,041 projects, 34,545 used Supabase, a popular hosted database. Of those, 24,664 sent database records as soon as the page loaded. That alone is not a flaw, because some data is meant to be public. But using only read requests with the same key the site gives every visitor, the researchers found email addresses in 10,560 tables across 7,403 sites. Examples include a German lifestyle forum with more than 80,000 profiles and 200,000 chat messages, a service provider with over 1,000 passport records, and 280,000 customer emails.
Nobody broke into Lovable or Supabase. As Trending Topics explains, the key a Supabase app sends to the browser is public by design, and protection depends on Row Level Security rules set on each table. If those rules are missing or too loose, anyone holding the key can read the table. Because AI-built apps follow similar patterns, one mistake can be searched for across thousands of sites. The problem is wider than Lovable: Trending Topics notes that UpGuard found readable tables on 16,326 of about 300,000 Supabase-linked domains.
There are real limits to the findings. Crissy Field says the 30% share (7,403 of 24,664) is not a confirmed leak rate, since some email addresses are public contact details or test data, and the Common Crawl sample is not representative of all Lovable apps. The firm could not establish whether any data was misused, and it is not naming the affected apps. Lovable's side is second-hand: Trending Topics, citing Cybernews, says the company sees a known, industry-wide Supabase configuration issue rather than a flaw in its AI, and points to its built-in security scan.
The practical point is about responsibility. Crissy Field notes that Lovable markets itself as needing no technical knowledge, while its terms put responsibility for apps and data on the customer. So a shop owner who built a booking app in a weekend may be the legal data controller for a database they cannot inspect. The researchers' advice is to have access permissions reviewed by someone qualified before an app handles real data, and again after big changes. For builders, that advice describes a service that many small businesses now need and few know to ask for.
Why it mattersAI app builders let non-developers ship real products, but the security setup still needs someone who understands it. Many small businesses are now holding customer data in databases nobody has checked.
👀 What to watch
Whether Lovable changes its defaults so new tables are locked by default, whether European data protection authorities act on the findings, and whether similar studies appear for other AI app builders.
3 opportunities from this story
Fixed-price security check for AI-built apps
Offer a short review for businesses that built an app with Lovable or a similar tool and connected it to Supabase. You check which tables a logged-out visitor can read, review the Row Level Security rules, and hand back a plain-language report with fixes. Buyers are founders and small firms that shipped fast and are now reading about leaks. The study gives you a clear, sourced reason for the conversation.
- Best for
- Developers and freelancers who know Supabase, Postgres and basic web security
- First step this week
- Build a small test app on a free Supabase project with one table left open, and write up exactly how you would find and fix it as your sample report.
Open the full playbook
Launch steps
- Write a checklist covering table policies, storage buckets, edge functions and exposed keys
- Create a two-page sample report from your own test app
- Get written permission from each client before testing anything
- Offer the first three reviews at a reduced fee in exchange for a testimonial
Tools
Risks
Testing an app without the owner's written permission can be illegal, so only review apps you are hired to review. A short check cannot promise an app is safe, so state the scope clearly in the report.
Launch-readiness package for vibe-coded apps
Many people can now build an app but cannot take it live responsibly. Sell a package that covers access rules, sign-in setup, backups, a privacy notice and a basic data map before launch. The buyer is a non-technical founder with a working prototype and real customers waiting. The study shows what skipping this step can look like.
- Best for
- Freelance developers and small agencies that already build or maintain web apps
- First step this week
- List the ten things you would check before letting a client's AI-built app take real customer data, and turn the list into a one-page offer.
Open the full playbook
Launch steps
- Define a fixed scope: access rules, auth, backups, privacy notice, error monitoring
- Price it as one package so buyers do not need to understand each item
- Run it on a friend's or your own prototype to time the work
- Post the checklist publicly and offer the package to people who reply
Tools
Risks
Scope can grow quickly if the prototype is messy, so cap the hours and price rework separately. You are not a lawyer, so point clients to a professional for privacy law questions such as GDPR or Kenya's Data Protection Act.
Plain-language guide and workshop on securing AI-built apps
Most people using AI app builders have never heard of Row Level Security. Create a short guide or a 90-minute live workshop that shows, step by step, how to check whether an app's data is open and how to lock it. Buyers are solo founders, students and small teams who want to fix it themselves. It is a low-cost product that can also feed your review service.
- Best for
- Developers who teach well, technical writers, and community organisers in builder groups
- First step this week
- Record a 10-minute screen video showing how to check one Supabase table's policies, and share it in one builder community to test interest.
Open the full playbook
Launch steps
- Outline five checks a non-developer can run in the Supabase dashboard
- Record each check on a demo project with no real data
- Package the videos with a printable checklist
- Run one free session, collect questions, then charge for the next
Tools
Risks
Platform screens and defaults change, so date the material and update it. A guide cannot cover every app, so say clearly that it is a first check and not a full security assessment.
Stanford's open-source Memorilla packs a document collection into 16 memory tokens for a frozen LLM
A team led from Stanford has released Memorilla, a 194-million-parameter module that compresses a collection of documents into 16 learned tokens for an unmodified Qwen3-8B model. The authors report 84.33% on the FactKG fact-checking test, ahead of putting every document in the prompt, but the method loses on some benchmarks.
Researchers from Stanford, the University of Pennsylvania, MIT and the University of Toronto have published Memorilla, a way to give a language model long-term memory without retraining it. The paper, first posted to arXiv on 15 October and revised on 3 November, lists Adibvafa Fallahpour as first author and Stanford professor Jure Leskovec as senior author. The code is public on GitHub under the Apache 2.0 licence, with trained checkpoints on Hugging Face, and the release was widely covered on 14 November. It needs no changes to the language model itself.
The design has three parts. A frozen embedding model, Qwen3-Embedding-4B, turns each document into a vector once and stores it. A small trainable module of 194,123,776 parameters then reads all the stored vectors alongside the user's question and produces exactly 16 'memory tokens'. These are placed in front of the question and passed to a frozen Qwen3-8B model. Only the small module is trained, in three stages that the GitHub page says take about 19 hours on four Nvidia H100 GPUs.
The reported results are strong on average. Across eight benchmarks the authors give Memorilla a mean score of 49.68, against 41.92 for the next-best method, xRAG. On FactKG it scores 84.33% versus 80.43% for full-context prompting, using 108 times fewer document tokens. On PersonaMem-v2, a test of remembering a user across conversations, it reaches 44.6% against 30.9% for the strongest baseline. With 512 documents on one H100, the paper reports an answer in 0.74 seconds versus 19.9 seconds for full context, about 27 times faster.
The caveats matter. These are the authors' own numbers and have not been independently reproduced at scale. Memorilla loses clearly on PubMedQA, with 59% against 69% for LLMLingua-2, a case where the source text is short and simply including it works better. A rival method, MemoRAG, leads on parts of NarrativeQA and PersonalizationV4 while using far more tokens. The authors say the module can lose fine-grained facts, was tested only with Qwen models, and was not evaluated on tasks such as coding. Running it needs one 80 GB GPU.
For builders, the idea is a possible alternative to stuffing long histories into prompts or running retrieval pipelines. If a chatbot can carry a customer's past conversations in 16 tokens, each reply costs far less and returns faster. That is most useful for personalisation and support tools that serve many users with long histories. It is less suited to work where exact wording matters, such as contracts or medical text. Because the code and weights are open, small teams can test the claim on their own data rather than take it on trust.
Why it mattersLong prompts are a major cost in AI products. A small add-on that compresses memory without retraining the model could make personalised assistants much cheaper to run, if the results hold up.
👀 What to watch
Independent reproductions, versions for model families other than Qwen, and whether hosted inference providers offer Memorilla-style memory as a service.
3 opportunities from this story
Benchmark report: Memorilla versus RAG on a client's own data
Teams running chatbots on retrieval pipelines want to know if a new memory method is worth a switch, but few have time to test it. Offer a short paid evaluation that runs Memorilla and the client's current setup on a sample of their real questions and reports accuracy, speed and cost side by side. Buyers are startups with support or knowledge assistants. The open code and checkpoints make this possible without training anything.
- Best for
- ML engineers and technical freelancers comfortable with GPUs and evaluation scripts
- First step this week
- Rent an 80 GB GPU for a few hours, run the repository's evaluation on one public benchmark, and confirm you can reproduce a published number.
Open the full playbook
Launch steps
- Reproduce one published result from the GitHub repository
- Build a small harness that scores answers from two systems on the same questions
- Run it on a public dataset and write a sample report
- Offer the evaluation to three teams with a retrieval-based product
Tools
Risks
GPU rental costs add up, so price them into the fee. Results on a client's data may be worse than the paper's, so sell an honest comparison and not a promised improvement.
Plain-English explainers on new AI memory methods
Founders and product managers hear about a new memory technique every month and cannot tell which matter. Write short, sourced briefs that explain what a method does, what it costs to run and where it fails, starting with Memorilla. Buyers are newsletters, agencies and product teams that need a clear summary. The paper's mixed results give you a real story to tell rather than a press release.
- Best for
- Technical writers, developer advocates and students who can read ML papers
- First step this week
- Write a 600-word brief on Memorilla covering what it is, the best result, the worst result and the hardware needed, and publish it.
Open the full playbook
Launch steps
- Read the paper's abstract, results table and limitations section
- Write one brief with every number linked to its source
- Publish it on LinkedIn or a blog as a sample
- Pitch a monthly 'AI memory and retrieval' briefing to two agencies or newsletters
Tools
Risks
Misreading a paper damages your credibility, so quote numbers exactly and say when they are self-reported. Paying readers are hard to win, so start with per-article work.
Personalised support assistant pilot with compressed memory
Support chatbots that remember each customer usually resend long histories with every message, which is slow and costly. Build a pilot for one business that uses a compressed-memory approach to recall past conversations, and measure reply time and cost against their current bot. Buyers are companies with repeat customers and high chat volume. Treat it as an experiment with a clear success test, not a finished product.
- Best for
- Small AI agencies and developers who already build chat assistants for clients
- First step this week
- Ask one existing client for a sample of anonymised chat histories and agree what 'better' would mean before you build anything.
Open the full playbook
Launch steps
- Agree a success metric such as reply time or cost per conversation
- Set up the open checkpoint on a rented GPU with anonymised data
- Compare answers against the client's current bot on the same questions
- Report the results honestly and recommend go or no-go
Tools
Risks
The method can lose fine details and is tested only on Qwen models, so it may not suit the client's use. The authors note stored embeddings can hold sensitive information, so handle customer data under a written agreement.
Hyundai Home Shopping airs a sales show made entirely by AI, with no host and no camera
South Korean TV retailer Hyundai Home Shopping said on 15 November that it will air a regular sales programme generated entirely by AI, starting with Jeju kiwis on 16 November. The company says a trial cut suppliers' production costs by more than 80% and sales reached 120% of target, figures it has not detailed.
Hyundai Home Shopping, one of South Korea's large TV shopping companies, announced on Sunday 15 November a system called AI-REAL that generates a full sales broadcast, both the voice and the pictures, with no human host and no camera shoot. The first regular programme is scheduled for Monday 16 November at 3:20 p.m. on its data shopping channel, PLUS#, selling Jeju Heart Gold kiwis, according to the Korea Herald. The company calls it a first for the country's home shopping industry, where AI had so far replaced single scenes or presenters.
The numbers come from a recent trial and from the company itself. Hyundai Home Shopping says suppliers' costs for preparing samples and producing supporting video fell by at least 80% compared with a normal production, and that sales reached 120% of target. It plans to air at least one AI-generated programme a month, starting with produce, meat and seafood, then moving into processed foods and household goods. The Seoul Economic Daily reports that those later categories are where smaller suppliers have shown the most interest.
The pitch is aimed at small suppliers. A traditional home shopping slot needs a studio, hosts, samples and filmed inserts, which is costly for a farm or a small food brand. The company says AI-REAL can also show things that are hard to film, such as the soil where a crop grows or how a product moves through distribution. Its head of broadcast production told the Korea Herald that trial viewers found the show about as natural as one with a human host, and that growing and cooking details helped them decide.
There is a lot the announcement does not say. No AI model or vendor is named. The trial's size, length and products are not disclosed, and '120% of target' means little without the target. The viewer feedback is the company's own account, not an independent measure. It is also unclear how humans review the content before it airs, which matters because a generated video of a farm or a product could show things that are not accurate. As of writing, the first regular broadcast had not yet aired, so its results are unknown.
The wider lesson is about cost. If a large retailer believes generated video can replace a studio shoot for selling fruit, small sellers everywhere will ask the same question about their own product videos. The tools to make short AI product clips are already available to freelancers. The gap is in doing it honestly: showing the real product, labelling generated footage where rules require it, and not inventing claims. Sellers on social platforms and marketplaces in markets such as Kenya rarely have a video budget, which makes low-cost, accurate product video a practical service to offer.
Why it mattersA mainstream retailer is using fully generated video to sell real products on air. If the cost savings are real, product video becomes affordable for sellers who could never pay for a shoot.
👀 What to watch
How the first broadcast sells, whether the company publishes real figures, and whether Korean regulators or rival channels respond on labelling of AI-generated sales content.
3 opportunities from this story
AI product videos for small food and farm sellers
Small sellers of produce, snacks and household goods rarely have video because a shoot costs too much. Offer short product clips made with AI video tools from the seller's own photos and facts, sized for WhatsApp status, TikTok and marketplace listings. Buyers are small brands, farmer cooperatives and market traders who sell online. The Hyundai announcement shows a large retailer betting on the same idea for its own suppliers.
- Best for
- Freelance video editors, content creators and social media managers
- First step this week
- Pick one local product, make a 20-second clip from real photos with an AI video tool, and show it to the seller.
Open the full playbook
Launch steps
- Collect real product photos and three true facts from the seller
- Generate and edit a short clip with a voiceover in the local language
- Get the seller to approve every claim before posting
- Offer a bundle of four clips a month after the first one
Tools
Risks
Generated footage can misrepresent the product, so base it on real photos and get written sign-off. Platform rules on labelling AI content vary, so check and label where required.
Product explainer scenes that are hard to film
Hyundai says its system can show soil conditions and distribution steps that cameras rarely capture. Offer the same for brands: short animated or generated scenes that explain where a product comes from and how it is made, based on facts the brand supplies. Buyers are exporters, agribusinesses and consumer brands that want to show their sourcing. The work sits between design and video and needs careful fact-checking.
- Best for
- Motion designers, small creative studios and agencies serving food or export brands
- First step this week
- Storyboard a 30-second 'farm to shelf' explainer for one product you know well, with a source for every step shown.
Open the full playbook
Launch steps
- Interview the brand to document each real step in sourcing and production
- Write a script where every scene matches a documented fact
- Generate scenes, then edit and label them as illustrations
- Deliver versions for web, social and trade-show screens
Tools
Risks
An illustrated scene can be taken as real footage, so label it clearly as an illustration. Claims about origin or farming methods must be true, or the brand could face consumer protection complaints.
Review checklist service for AI-generated sales content
As more sellers use generated video, someone needs to check it before it goes out. Offer a quick review that compares each clip against the product's real specifications, the platform's advertising rules and local consumer law basics. Buyers are agencies and brands producing AI content in volume. The Hyundai announcement did not describe its human review step, which is exactly the gap this service fills.
- Best for
- Compliance-minded marketers, former ad reviewers, and paralegals with marketing experience
- First step this week
- Draft a one-page checklist covering product accuracy, price claims, AI labelling and required disclaimers, and test it on five AI-made ads you find online.
Open the full playbook
Launch steps
- Collect the ad and AI-content rules of the main platforms your clients use
- Build a checklist and a simple pass, fix or reject report format
- Test it on public examples and refine it
- Offer a trial month to one agency producing AI video
Tools
Risks
You are giving a review, not legal advice, so say so and refer legal questions to a lawyer. Rules change often, so date your checklist and update it.
Get these as a PDF every 3 days
Free. One email every 3 days. Unsubscribe any time.