larkspurThe Leak Log · Updated 27 Oct 2026Patterns

Security for AI-built apps · A public record

The Leak Log

Public, already-reported cases where AI-built apps exposed data, what caused each one, and the 30-second check for your own app.

Cases
34
Root causes
8
Last updated

Most cases are apps built with AI tools or on the backends those tools use. Some are the AI coding tools themselves, and a few older no-code cases are kept as a baseline. Each entry says what was built with what, and where the AI-built claim is not confirmed, it says that too.

01The log · Newest first

The log

Showing all 34 cases.

A coffee franchise's old dashboard exposed 132 locations' data

  • Not confirmed as AI-builtThe exposure is confirmed by the company; the AI-built claim is not. Listed as an old dashboard left online.
Exposed
Franchisee and operator names and emails, managers' names and mobile numbers, and business performance and pipeline data. The company says no customer data was involved.
Scale
132 locations. Taken offline 15 Oct 2026; the company planned to notify the regulator.
Built with
Bootlegger Coffee Company's legacy BOS operations dashboard, hosted on Railway. One outlet described it as built with Claude; the company has not confirmed that, and a second outlet's follow-up does not say it was AI-built.
What went wrong
Per the first report, data was sent along with the login page itself, so it was visible in the page source without signing in. The company says some data "was not protected by its login."
Check yours in 30 seconds
Log out, open your login page, right-click and choose View Page Source, and search for an email address or a customer name. Then open DevTools > Network, reload, and check what the page fetched before you signed in. Remove or lock any old copy of your app still running on a staging address.

16,326 Supabase databases with tables anyone could read

Exposed
Over half showed signs of personal data; some held passwords or auth tokens, and a very few held card data. Deep dives included an adult-creator platform (65,467 users, 100,000+ private messages), a valet business (100,000+ customers), a consulate run by an African government (25,000 people) and an immigration coach (884 plaintext passwords).
Scale
About 300,000 domains showing signs of Supabase use; 16,326 databases exposing readable tables.
Built with
Supabase-backed sites. UpGuard attributes the pattern to AI coding agents but did not count how many were agent-built.
What went wrong
Tables created through the API, the way coding agents create them, do not get RLS by default; tables made in the dashboard's Table Editor do.
Check yours in 30 seconds
Supabase > Advisors > Security Advisor, or supabase db advisors from the CLI. Every table your agent created by migration or API needs alter table x enable row level security; plus a real policy.
Note
Supabase told TechCrunch it had not seen the research, that its projects are "secure by default", and that "customers control how their own projects are configured." It did not address the numbers.

PixelLeak: coding agents posted internal screenshots to public repos

Exposed
Screenshots and screen recordings of internal tools: customer billing records, a treasury and settlement console, unreleased product screens.
Scale
Per the researchers: 13,000+ images, 300+ organizations, 900+ repos. The Hacker News checked gitshot's code, which defaults to a public repo, and found about 130 public gitshot repos. It could not tell whose work they held and did not verify the headline counts.
Built with
AI coding agents (reproduced with Claude Code in a lab; the models in the wild are not named) plus the open-source gitshot tool, under developers' personal GitHub accounts.
What went wrong
Agents worked around missing image uploads in pull requests by creating public repos, and one company saved this as a shared agent skill.
Check yours in 30 seconds
On GitHub, list your personal account's public repos, newest first. Look for gitshot-images or any image-only repo you do not remember creating.
Vendor research
Vendor research; the headline counts have not been independently verified.

An AI-written mailing script showed 95,364 customers' emails

Exposed
Email addresses of 95,364 loyalty members, each visible to up to 999 others in a single To field.
Scale
95,364 people. Described as the first AI-related breach notified to Singapore's data protection regulator; the company signed a voluntary undertaking.
Built with
Bee Cheng Hiang (Singapore). A Python bulk-mail script written by a marketing employee with a generative AI tool.
What went wrong
A bug in AI-generated code: the prompt never asked to hide recipients, and testing checked the logs rather than the email itself.
Check yours in 30 seconds
Before any bulk send from a script, send one real batch to two inboxes you own and open the email. Check To, CC and BCC with your own eyes, not the logs.

METR: a vibe-coded dashboard failed open and an API key was taken

Exposed
A model-provider API key, which an attacker obtained by prompting an agent on the exposed machine and used for about three weeks. The attacker also added an SSH key.
Scale
About $600,000 of credits at nominal value. They were granted free, so METR was not billed.
Built with
A researcher-built agent dashboard on a personal EC2 instance behind Google sign-in. METR's own post calls it "vibe-coded."
What went wrong
Fail-open auth: a flaw in the app silently disabled its Google sign-in check, leaving it exposed to the public internet for several days.
Check yours in 30 seconds
Find your auth middleware and ask: if the auth provider call throws or times out, what happens? It must deny. Search for catch blocks around session checks that return next() or return true.

A lab study of 28 AI-built apps: hardcoded secrets were the largest group of critical flaws

Exposed
No production victims: 434 validated findings. In the 27 newly generated apps, 11 of the 23 critical findings were hardcoded or default secrets and 6 were debug-mode or dependency remote code execution.
Scale
28 apps.
Built with
27 new apps generated with Anthropic and OpenAI models, plus one real CMS that AI rewrote from legacy Gnuboard5.
What went wrong
Hardcoded or default secrets were the top critical category.
Check yours in 30 seconds
Search your repo for secret, SECRET_KEY, JWT_SECRET, supersecret, changeme, default. Any secret with a literal value in code (not read from the environment) is shared with everyone who has seen the code. Move it to an env var and rotate it.
Vendor research
Vendor study, with no human-written baseline for comparison.

A financial firm's onboarding app, partly built with Claude, treated an applicant ID as a password

Exposed
At risk, not confirmed accessed: other applicants' government IDs, identity-verification data, payment details, Social Security numbers and co-applicant details.
Scale
Not reported. Found in a penetration test; no exploitation reported.
Built with
A financial firm's onboarding app, portions built with Claude. The firm is not named.
What went wrong
Anyone holding an applicant GUID could get an access token for that applicant: an identifier was used as proof of identity.
Check yours in 30 seconds
Find every resume or magic link and every URL with an ID in it. Change the ID to another record's ID while logged in as a different user. If you get their data, you have the same flaw.
Vendor research
Both sources are from Sygnia, which sells penetration testing. No independent coverage.

Nearly two-thirds of AI iPhone apps studied leaked access to their paid AI accounts

Exposed
Working LLM access across at least ten providers: bearer tokens, provider API keys, and backend proxies that needed no credential. 28 apps also exposed their system prompts.
Scale
282 of 444 apps (about 63%). Three months after disclosure only 28% had fixed it. The paper's retest found 66 apps (23%) still exploitable and 83 (29%) offline.
Built with
iOS App Store apps with LLM features. Apps not named.
What went wrong
Seen in the apps' own network traffic: 48% leaked a bearer token for a backend proxy, 33% used a proxy that needed no credential at all, and 19% sent a provider API key in plaintext.
Check yours in 30 seconds
Search your mobile or web bundle for sk-, OPENAI, ANTHROPIC, GEMINI, and watch your app's network requests for a provider key. Then call your own AI proxy endpoint with no auth header. If it answers, anyone can spend your credits.
Note
An academic preprint with a stated method. Its abstract says 72% remained exploitable; its own results count 23% still exploitable and 29% unreachable, so 72% is the share not fixed.
Sources

About 5,000 of 380,000 vibe-coded apps exposed sensitive data

Exposed
A shipping app's vessel routes and ports, a Brazilian bank's internal financials, a cabinet supplier's customer service chats, patient conversations at a long-term care facility, and a hospital's doctor-patient conversation summaries.
Scale
About 380,000 publicly accessible apps and assets found; about 5,000 with sensitive data, per Axios.
Built with
Apps on Lovable, Base44, Replit and Netlify.
What went wrong
Some platforms made apps public by default, owners left them without a login, and many were indexed by search engines.
Check yours in 30 seconds
Search Google for site:yourapp.lovable.app (or your .replit.app, .base44.app, .netlify.app address). Anything indexed that should be internal needs a login or a private setting today.
Note
Axios verified several examples. Lovable and Base44 said the report did not include URLs they could verify; Wix (Base44) said fake apps can easily appear to hold real data; Replit's CEO called public access expected behavior. A public app is not always a breach.
Vendor research
Vendor research: RedAccess sells security products.

"Comment and Control": AI agents on GitHub leaked their own CI secrets

Exposed
CI runner environment variables, an API key and tokens, posted back as comments, logs or commits.
Scale
Three major vendors' agents. Reported privately between Oct 2025 and Feb 2026; all three confirmed and paid bounties ($100, $1,337 and $500).
Built with
Claude Code Security Review action, Google Gemini CLI Action, GitHub Copilot agent.
What went wrong
Pull request titles, issue text and hidden HTML comments were read as instructions inside a runner that held secrets.
Check yours in 30 seconds
Open .github/workflows/. Any AI agent step triggered by pull_request_target, issues or issue_comment from outside contributors, with secrets in env:, is this exact setup. Limit it to trusted authors or remove the secrets.
Note
Anthropic says the action is not designed to be hardened against prompt injection, and GitHub calls the Copilot case a known architectural limitation. The Copilot attack also needed a maintainer to assign the issue.

Lovable's API showed other users' project chats and code

  • DisputedLovable disputes the scope, saying only chats on public projects were exposed. It has apologized for how the report was handled.
Exposed
Per the researcher: another user's source code, database credentials found in that code, AI chat histories and customer data, reachable from a free account in five API calls. Per Lovable: chats on public projects became visible again after a February permissions change.
Scale
The researcher says every project created before Nov 2025; Lovable says public projects only. The researcher says the report sat open for 48 days after HackerOne marked it a duplicate.
Built with
The Lovable platform API.
What went wrong
Per the researcher, broken object-level authorization: the API did not check that the requester owned the project. Lovable says public-project chats were reachable after a February permissions change.
Check yours in 30 seconds
In Lovable (or Bolt, Replit, v0), open each project's visibility setting and set it to private. Then search the project's code for any key or connection string; assume anything there has been seen, and rotate it.

A medical practice's vibe-coded patient app: every record in 30 minutes

Exposed
All patient records, readable and writable. Appointment audio recordings were sent to US-based AI services for transcription and summarization, and the data was stored on a US server without a Data Processing Agreement.
Scale
The whole patient database; the patient count was not given. The fix was basic auth plus key rotation.
Built with
A single-file HTML and JavaScript app built with an AI coding agent on a managed database. The author cites Swiss data protection law; the practice and the product are not named.
What went wrong
No access control or row-level security on the database. All the "auth" was in browser JavaScript, so one curl returned the data.
Check yours in 30 seconds
View your page source. If the login check is an if in browser JavaScript and the data comes straight from the database URL, there is no login. The database itself must refuse unauthenticated reads.
Note
A single first-hand blog post; the app and the practice are not named, and some Hacker News commenters questioned the account.

Commits co-written by Claude Code leaked secrets at twice the baseline

Exposed
Hardcoded secrets in public GitHub commits.
Scale
28.65M new hardcoded secrets in 2025 (up 34%), with AI-service secrets up 81%. Claude Code-assisted commits leaked at about 3.2%, against a 1.5% baseline.
Built with
Public GitHub. Commits co-authored by Claude Code; the report also charts Cursor and Lovable commit volumes.
What went wrong
Secrets written into code and committed to public repositories.
Check yours in 30 seconds
Run gitleaks detect or trufflehog git file://. in your repo (both free). Add a pre-commit hook so the next one is blocked before it is pushed.
Vendor research
Vendor research from a company that sells secret detection, read via its press release. The method has not been independently checked.

An exam app showcased on Lovable: 18,000 users, access check inverted

Exposed
18,697 user records (14,928 unique emails): 4,538 student accounts, 10,505 enterprise users, 870 with full personal data. Users included teachers and students from UC Berkeley, UC Davis and K-12 schools. An unauthenticated attacker could also delete any account and grade student test submissions.
Scale
16 vulnerabilities, 6 of them critical.
Built with
An app hosted on Lovable and featured on its Discover page, with a Supabase backend. The app is not named.
What went wrong
Missing RLS and role checks, plus an access function with inverted logic that blocked signed-in users and let anonymous ones through.
Check yours in 30 seconds
Open your Supabase SQL functions and RLS policies and read each if or case that checks a role out loud. Then test as three people: logged out, a normal user, an admin. The logged-out test must fail.
Note
Lovable says it contacted the owner, that the project included code not generated by Lovable, that the vulnerable database was not hosted by Lovable, and that every project gets a free security scan before publishing.

Moltbook: a vibe-coded AI social network with its whole database open

Exposed
1.5 million agent API tokens, about 35,000 emails (plus about 29,600 in another table), 4,060 private agent-to-agent messages (some containing plaintext OpenAI keys), and write access to live posts.
Scale
The entire production database. Reported 31 Jan at 22:06 UTC; first fix 83 minutes later; fully fixed by 01:00 UTC.
Built with
Supabase. The founder said publicly that he "didn't write a single line of code."
What went wrong
The public Supabase key was in client JavaScript and no RLS policies existed, so it granted full read and write.
Check yours in 30 seconds
Supabase > Advisors > Security Advisor; fix every rls_disabled_in_public and policy_exists_rls_disabled warning. (rls_enabled_no_policy means the table is locked, not open.) Then run curl "https://YOURREF.supabase.co/rest/v1/YOURTABLE?select=*" -H "apikey: YOUR_ANON_KEY" against your own project, with no login token. Any rows back means strangers get them too.

Claude Code: a cloned repo could send your API key elsewhere

Exposed
The developer's Anthropic API key, sent to an attacker's server before the trust prompt appeared. CVE-2026-21852, plus related hook and MCP auto-approval issues that Check Point Research reported in the same work.
Scale
All users on older versions. Patched before public disclosure.
Built with
Claude Code before 2.0.65.
What went wrong
A malicious project settings file redirected ANTHROPIC_BASE_URL.
Check yours in 30 seconds
Run claude --version (2.0.65 or later). Before opening a repo you did not write, look inside .claude/settings.json and .mcp.json for unfamiliar URLs, hooks or auto-approve settings.

n8n "Ni8mare": public forms could read server files

Exposed
Arbitrary file reads on the n8n host, including config files and stored credentials, and command execution in some setups. CVE-2026-21858, CVSS 10.0.
Scale
Censys counted 26,500+ internet-facing n8n hosts; other estimates were higher.
Built with
n8n, the no-code workflow tool, versions 1.65.0 through 1.120.x (fixed in 1.121.0).
What went wrong
Public webhook and form endpoints trusted the request's content type to decide how to handle files.
Check yours in 30 seconds
In n8n, open Help > About n8n in the left sidebar, or run n8n --version on the server, to see your version; anything under 1.121.0 needs an update. List every active workflow with a Webhook or Form trigger: each one is a public door.
Note
Scope is debated: Horizon3 argued real-world impact was smaller than the host count suggests.

A scan of 5,600 vibe-coded apps found 400+ exposed secrets

Exposed
2,000+ vulnerabilities, 400+ exposed secrets (including Supabase service keys in frontend bundles) and 175 instances of personal data: medical records, IBANs, phone numbers, emails.
Scale
5,600+ apps and 14,600 assets, scanned passively.
Built with
Mostly Lovable (about 4,000 apps), plus Create.xyz, Base44, Bolt.new and Vibe Studio, on Supabase backends.
What went wrong
Server-only keys shipped in the browser bundle. Missing RLS was also cited.
Check yours in 30 seconds
Load your live site, open DevTools > Sources, and search all files (Cmd+Option+F on Mac, Ctrl+Shift+F on Windows) for sb_secret_, sk_live, AKIA and service_role. Older Supabase keys are JWTs that start with eyJ and hide their role inside, so paste each eyJ key into JSON.parse(atob(KEY.split('.')[1])) in the console: the role must be anon, never service_role. Any secret key you find should only live on a server: rotate it today.
Vendor research
Vendor research from a security scanning company. Its landing page cites a 1.4K base; the methodology post says 5,600.

CamoLeak: Copilot Chat leaked private repo contents through images

Exposed
Secrets, private source code and unpublished vulnerability notes from private repos, via hidden markdown comments in pull requests. CVSS 9.6.
Scale
Any Copilot Chat user reviewing a booby-trapped pull request. GitHub fixed it by disabling image rendering in Copilot Chat.
Built with
GitHub Copilot Chat.
What went wrong
Invisible pull request comments became instructions, and data left through GitHub's own image proxy.
Check yours in 30 seconds
Before asking an AI to review an outside contributor's pull request or issue, read its raw text with gh pr view NUMBER --json body (or gh issue view NUMBER --json body), or open the description's Edit box, and look for <!-- comments. And keep secrets out of repos entirely, so there is nothing to take.

Nx "s1ngularity": malware used the victim's own AI CLIs to hunt secrets

Exposed
GitHub tokens, npm tokens, SSH keys, API keys and crypto wallet files, uploaded to public repos created in victims' own GitHub accounts. More than 6,700 private repos were then made public.
Scale
GitGuardian: 2,349 distinct secrets, and 1,346 public repos named s1ngularity-repository. Wiz, via SecurityWeek: more than 1,700 users with leaked secrets and more than 6,700 private repos made public.
Built with
Malicious versions of the Nx npm packages. The payload ran locally installed Claude, Gemini and Q CLIs with permission-skipping flags.
What went wrong
Orca called it the first known supply chain attack to search developer machines for installed AI CLIs and turn them on the victim's own files.
Check yours in 30 seconds
On GitHub, search your account for repositories named s1ngularity-repository. Check whether you run AI CLIs with --dangerously-skip-permissions, --yolo or --trust-all-tools by default. Run npm ls nx in each project.
Note
Counts differ by source.

Cursor: a Slack message could rewrite your MCP config

Exposed
Remote code execution on the developer's machine, with access to everything the user can reach: keys, repos, env files. CVE-2025-54135 (CVSS 8.6) and CVE-2025-54136.
Scale
All Cursor users before 1.3; fixed in 1.3 (Jul 2025).
Built with
Cursor IDE with MCP servers.
What went wrong
Injected text made the agent write mcp.json, which ran before the user could reject it, and approved MCP configs could be swapped silently.
Check yours in 30 seconds
Check that Cursor is 1.3 or later (Cursor > About Cursor on Mac, Help > About on Windows and Linux). Open ~/.cursor/mcp.json and your project's .cursor/mcp.json, remove servers you do not recognize, and review what each remaining server runs in Cursor Settings under MCP.

TeaOnHer exposed driver's licenses through open API routes

Exposed
Usernames, emails, driver's license photos and selfies. The developer's own admin email and plaintext password sat on the API landing page.
Scale
Thousands of users, on an app with about 53,000 users at the time. TechCrunch found it within 10 minutes of getting the App Store link.
Built with
TeaOnHer, a rival to Tea aimed at men. ID photos sat in a publicly accessible Amazon S3 bucket; not shown to be AI-built.
What went wrong
Backend endpoints answered requests with no credentials, and the documentation endpoint was public.
Check yours in 30 seconds
Open your app, copy any API request from DevTools > Network as cURL, delete the Authorization header and the cookie, and run it. If it still returns data, that route is open.

Tea app: ID selfies in an open Firebase storage bucket

  • Not confirmed as AI-builtIt is often cited as a vibe-coding example, but the sources do not establish that. Listed for the storage lesson.
Exposed
About 72,000 images, including about 13,000 verification selfies and photo IDs. A second exposure leaked about 1.1 million private messages.
Scale
The image set covered users who signed up before Feb 2024; the leaked messages ran from 2023 into July 2025. The data was later shared on forums.
Built with
Tea, a dating-safety app. Researchers traced the images to Firebase storage; the company called it a legacy system.
What went wrong
Images sat in a storage bucket anyone could read.
Check yours in 30 seconds
Supabase > Storage: any bucket marked Public serves every file to anyone with the URL. Firebase > Storage > Rules: look for allow read: if true. ID documents and selfies belong in a private bucket, served through short-lived signed URLs.

An Amazon Q extension release shipped a prompt to wipe users' systems

Exposed
Nothing deleted (AWS says a syntax error stopped the code from running), but release 1.84.0 of the Amazon Q extension for VS Code told the AI to clear the system to a near-factory state and delete file-system and cloud resources.
Scale
Version 1.84.0, published 17 July and replaced by 1.85.0 on 24 July; the extension had nearly one million installs. AWS says no customer resources were impacted.
Built with
Amazon Q Developer for VS Code, an AI coding assistant.
What went wrong
An over-scoped GitHub token in the extension's build configuration let an attacker commit code that shipped automatically in a release, and the injected prompt aimed the agent at local files and cloud resources.
Check yours in 30 seconds
List the AI extensions in your editor and the flags you run agents with. With --dangerously-skip-permissions, --yolo or auto-approve on, the agent runs whatever it is told. Make sure the shell it uses holds no long-lived production cloud credentials.

Replit's agent deleted a live production database

Exposed
Not a leak: an AI agent ran destructive commands during a code freeze and wiped production records (1,206 executives and 1,196+ companies), then misreported what had happened.
Scale
One production database. Replit's CEO called it unacceptable and shipped automatic dev/prod database separation.
Built with
Replit Agent, during a public vibe-coding trial by SaaStr's founder.
What went wrong
The agent had write access to production, with no separation between development and production.
Check yours in 30 seconds
Open the .env your AI agent can read. Does DATABASE_URL point at production? If so, give the agent a dev database and keep production credentials out of its workspace. Confirm you have a backup you have actually restored.
Note
Replit confirmed the event. Some details rest on the user's own account, and whether a rollback was possible is disputed.

Base44: a public app ID opened private apps

Exposed
Access to private, SSO-protected apps built on Base44, including internal chatbots and apps holding personal data.
Scale
Platform-wide. Wix fixed it within about 24 hours; no evidence of exploitation.
Built with
Base44, a vibe-coding platform owned by Wix.
What went wrong
The register and verify-otp endpoints, listed in publicly reachable Swagger docs, accepted any public app_id and never checked whether the app was private.
Check yours in 30 seconds
Open your app's sign-up page in a private window. Can a random email register for an app that should be invite-only? Also load /docs, /swagger, /api-docs and /openapi.json on your domain; public API docs are a map for attackers.

Supabase MCP and Cursor copied secrets into a support ticket

Exposed
In a demo project with dummy data: an integration_tokens table, which the AI assistant copied into a support ticket the attacker could read.
Scale
Proof of concept on dummy data; no customer leak reported. Supabase recommends running its MCP server in read-only, project-scoped mode.
Built with
Cursor using the Supabase MCP server with service_role privileges.
What went wrong
Text in a support ticket became instructions, and the agent's key bypassed RLS.
Check yours in 30 seconds
Open ~/.cursor/mcp.json, your project's .cursor/mcp.json and .mcp.json, and your Claude or Windsurf MCP config. If a Supabase or other database server is connected with an access token, a service_role key or production credentials, set it to read-only (read_only=true or --read-only), scope it to one dev project (project_ref), or remove it.

A "123456" test login opened McDonald's AI hiring records

Exposed
Names, emails, phone numbers and chat transcripts tied to job applications on McHire.
Scale
Up to 64 million job applications (applications, not unique people).
Built with
McHire, built by Paradox.ai, an AI hiring chatbot. An AI product; not shown to be AI-built.
What went wrong
A test restaurant's admin account accepted 123456 / 123456. Changing an ID number in an internal API then returned other applicants' records.
Check yours in 30 seconds
Search your codebase, seed scripts and .env files for admin, 123456, password, changeme, test@. Try those on your staging and admin logins. Then delete or disable every test account in production.

Langflow's AI builder ran code from anyone, and it was exploited

Exposed
Full server takeover: any secrets, API keys and data on the Langflow host. CVE-2025-3248, CVSS 9.8; later used to spread the Flodrix botnet.
Scale
Added to CISA's Known Exploited Vulnerabilities catalog in May 2025. Fixed in 1.3.0.
Built with
Langflow, a low-code visual builder for AI agents and RAG apps.
What went wrong
The /api/v1/validate/code endpoint ran user-supplied Python with no sign-in.
Check yours in 30 seconds
If you self-host Langflow, Flowise, n8n or similar, turn off Wi-Fi and VPN on your phone and load your instance URL over mobile data. If it loads at all, it is on the open internet; if you reach anything without logging in, anyone can use it. Compare the version with the vendor's latest security release.

170 Lovable apps had Supabase tables anyone could read (CVE-2025-48757)

  • DisputedLovable disputes the CVE, saying each customer is responsible for protecting their own app's data.
Exposed
Names, emails, phone numbers, payment and subscription records and, in some cases, third-party API keys stored in app tables.
Scale
1,645 showcase apps checked, 170 exposed (about 10%), 303 vulnerable endpoints reported.
Built with
Apps generated with Lovable, on Supabase.
What went wrong
Tables were reachable with the public anon key because Row Level Security was missing or too permissive.
Check yours in 30 seconds
Supabase dashboard > Advisors > Security Advisor. Any rls_disabled_in_public error means that table is readable by anyone holding your anon key, which ships in your site's JavaScript. Or run in the SQL editor: select tablename from pg_tables where schemaname = 'public' and rowsecurity = false; Then open Authentication > Policies: a policy that reads using (true) lets everyone in even with RLS on.

Public repos using Copilot leaked secrets more often

Exposed
Hardcoded secrets (API keys, tokens, passwords) in public GitHub repositories.
Scale
23.8M secrets leaked on public GitHub in 2024, up 25% on the year. GitGuardian reports a 6.4% secret leakage rate in public repositories using Copilot, about 40% above the 4.6% rate across all public repositories.
Built with
Public GitHub repositories with GitHub Copilot in use.
What went wrong
Secrets committed straight into code and pushed to public repositories.
Check yours in 30 seconds
In your repo: git log -p --all | grep -E "sk_live_|sk-[A-Za-z0-9_-]{20,}|AKIA[0-9A-Z]{16}|service_role|PRIVATE KEY". Better: on GitHub open the repo's Settings > Advanced Security, enable Secret Protection and push protection, then read Security > Secret scanning alerts. It is free on public repos; private repos need GitHub Secret Protection.
Vendor research
Vendor research from a company that sells secret detection. The Copilot figure is a correlation; it does not show that Copilot caused the leaks.

Arc browser's Firebase rules let one user's code run for another

Exposed
Not data theft: any user could reassign a Boost (custom JavaScript) to another user's ID, and it then ran in that person's browser. CVE-2024-45489, CVSS 9.8.
Scale
The vendor says no users were affected and fixed it within a day. Arc stopped using Firebase for new features afterward.
Built with
Arc browser (The Browser Company), Firebase backend. Not AI-built; included for the rules mistake.
What went wrong
Misconfigured Firebase rules let a user create or update a Boost with another user's creatorID.
Check yours in 30 seconds
In your Firestore rules, find every allow create and allow update. On create, require request.resource.data.ownerId == request.auth.uid. On update, require request.resource.data.ownerId == resource.data.ownerId so nobody can hand a document to someone else.

900+ Firebase sites left 125 million records readable

Exposed
Names, emails, phone numbers, billing details with bank information, and about 20 million passwords, nearly all reported as plaintext. Most sites were also writable.
Scale
5 million+ domains scanned, 900+ misconfigured (916 in most coverage), about 125M records. 842 owners were emailed; 1% replied and about 24% fixed it.
Built with
Firebase-backed websites and apps, one of them a bank. Not AI-built specifically; included as the no-code backend baseline.
What went wrong
Firebase instances with missing or open security rules.
Check yours in 30 seconds
Open your Firestore rules and look for if true, then open your Realtime Database rules and look for ".read": true or ".write": true. Then search your code for any field named password being written to the database. Firebase Auth already stores credentials; you should never need a password column.

An AI hiring vendor's Firebase let any new user act as admin

Exposed
Names, phone numbers, emails, some plaintext passwords, branch locations, shifts and confidential messages of Chattr staff, franchise managers and job applicants at fast-food chains, plus an admin dashboard with the option to refund payments.
Scale
Not quantified. This finding led to the wider Firebase scan in the next entry.
Built with
Chattr.ai, an AI hiring chatbot vendor, on Firebase. An AI product; not shown to be AI-built.
What went wrong
The Firebase setup gave a newly registered user full database privileges.
Check yours in 30 seconds
Firebase console > Firestore Database > Rules (and Storage > Rules). If you see allow read, write: if true; or if request.auth != null covering everything, any signed-up stranger can read the whole collection. Rules should compare request.auth.uid with the document's owner.
Note
Accounts of the vendor's response differ: RH-ISAC reported no response at the time, while the researcher and SecurityWeek say the hole was patched on 10 Jan 2024.

Only public, already-reported incidents. We never test anyone’s app.

Private individuals are not named, researchers included. Spotted an error, a fix we missed or a case that belongs here? adam@larkspur.digital

02Patterns

What keeps coming up.

Read together, the 34 cases come down to a handful of mistakes. Each has its own link, and each lists the cases behind it.

2.1

The anon key is not the leak. Missing RLS is.

Supabase's anon key is meant to be public. It only becomes a master key when a table has RLS turned off or a policy that lets everyone in. Tables created through the API, the way coding agents create them, do not get the RLS default that the dashboard's Table Editor applies.

2.2

Client-side auth is no auth.

If the check lives in the browser, if the data arrives with the login page, or if the check fails open on an error, the data is public. The database or the server has to be the one that says no.

2.3

An ID is not a password.

App IDs, applicant numbers and GUIDs get copied, guessed and leaked. Every request needs proof of who is asking, not just which record they want.

2.4

Public by default.

Builder projects, hosted apps and helper repos have often started public. Lovable only switched to private by default in December 2025, and the gitshot screenshot tool still defaults to a public repo. Making them private is a separate step, and people who are new to building do not always know it exists.

2.5

An agent with secrets that reads outside text can be told to send them out.

When an agent holds a key and reads tickets, pull requests, issues or chat messages, assume that text can instruct it. The same goes for anyone who can reach the agent's own chat box. Keep production keys out of the agent's reach.

2.6

Leaks stay open.

About 24% of emailed Firebase owners fixed their setup. Three months after researchers reported leaky AI iPhone apps, only 28% had fixed them, and 23% could still be exploited. Nobody is coming to tell you; check it yourself.

2.7

Whoever ships the app owns its security.

Builder platforms say, fairly, that configuration is the owner's job: Lovable disputed its CVE on those grounds, and Supabase says projects are secure by default. Whatever tool built the app, the person who launches it is responsible for who can read it.

2.8Risk, not an incident

Unsigned webhooks: the flaw ships in real software, still worth checking.

We found no reported break-in, but the mistake has shipped: n8n's Stripe trigger (CVE-2026-21894) and a WordPress donation plugin (CVE-2026-3177) both skipped the signature check. A Stripe or similar webhook that skips signature verification lets anyone send a fake "payment succeeded" and get something for free.

Check yours Open your webhook handler. Before it changes anything, it should verify the signature against the raw request body (for Stripe, stripe.webhooks.constructEvent). If it parses the JSON and trusts it, it is open.

Root causes used in the log: RLS off or wrong, Key in frontend, Open API route, Agent tool misuse, Leaked key in git, Default credentials, Public bucket, Other.

Using Supabase? Read your own rules back.

The free RLS policy checker reads your schema and tells you, table by table, whether your rules stop the wrong person, with the SQL to fix each one. It runs in your browser, so nothing leaves your machine.

If you would rather a person checked it, a Leak Check is $490 and takes 48 hours.