We built a notes app with an AI coding tool and left in the usual mistakes: service-role key in the browser, RLS as a TODO, a Stripe webhook that believes anyone.
Then we ran our ship check on it. 16 launch blockers. 38 seconds:
The grep commands for a clone and for your own agent configs, what it steals, and how to prove the old tokens are dead: https://t.co/2TZt4W92Hp Ship Check Kit ($29) runs the rotate-and-verify checks on your app: https://t.co/2bs7DZ0Ood
A repo's agent instruction file can tell your coding agent to run a script it must not read. Bitdefender found 22 GitHub links doing it with a "sealed" scripts/setup.js, plus Claude Code session-start hooks that run it once you trust the folder. Read them before opening a clone.
The gh/git commands to find the probe PR, the reverted commits and run-number gaps, and how to prove old AI keys are dead: https://t.co/LC2WPS0OMd Ship Check Kit ($29) runs the rotate-and-verify checks on your app: https://t.co/2bs7DZ0Ood
wfx-loot reverts its commit within a minute, then deletes the workflow run and its logs. Per OX, the most reliable trace left is a missing run number. What it collects includes OpenAI, Anthropic, Gemini, Hugging Face, Cursor and Copilot keys. Check your Actions history.
Who's affected (only people Cursor contacted), the check for the Oct 29 extension, and how to prove old keys are dead: https://t.co/eRga0wLmFj
Ship Check Kit ($29) runs rotate-and-verify checks on your own app: https://t.co/U6O0MVuMwv
Cursor's Nov 13 breach update, in order:
1. A key it gave Fireworks leaked
2. Old agent traces got copied (<0.05% of users)
3. Some held raw credentials
4. Those reached keys users sent support
5. A tampered Claude Code extension hit its marketplace
Ever paste keys into agents?
These are 5 of the 44 checks in Ship Check Kit. Your coding agent runs them on your repo and live app, then hands you a pass/fail report. $29.
https://t.co/U6O0MVvkm3
5 checks to run before your launch post:
1. Log out, call your API routes
2. Log in as user B, swap in user A's IDs
3. Query every table with the public anon key
4. Send your Stripe webhook an unsigned event
5. Find your AI spend's hard ceiling
Which one have you never tried?
teai's report, quoted, plus the checks to run through your own AI's tool: https://t.co/Fi3kDFqzDw
Ship Check Kit has your coding agent run 44 launch and security checks before you ship. $29. https://t.co/U6O0MVvkm3
teai gave its chat AI a code_execute tool. It ran as root, on the service's own machine.
Ask the AI, read the env vars: ~100 production secrets, for two months. Then 75,964 phishing emails from six of its domains.
The only guard was the model.
What can your agent's tools read?
Claim steps, refund terms by provider, and the one-line 401 check: https://t.co/iaCOz6h7sp
Ship Check Kit has your coding agent run 44 launch and security checks, including proof each old key is dead. $29. https://t.co/U6O0MVvkm3
Railway's incident report puts spend on the stolen AI keys at ~$740K. Railway says it will pay whatever the providers won't refund.
Claims are due Wed Nov 18.
A claim doesn't revoke anything, though. Only the provider that issued the key can.
Does your old key return 401 yet?
The three checks that would have stopped it: https://t.co/pOMCQVEScR
Ship Check Kit has your coding agent run 44 launch and security checks, with a fix for every failure. $29. https://t.co/U6O0MVvkm3
A vibe-coded agent dashboard's login failed open. Nobody noticed.
METR, disclosed Aug 31: the attacker asked the agent for its API key, added an SSH key, and ran on its credits for three weeks. ~$600K worth. Theirs were donated. Yours won't be.
Does your login fail closed?
What it means and what to rotate: https://t.co/ME49uuGBSV
Ship Check Kit has your coding agent run 44 launch and security checks, with a fix for every failure. $29. https://t.co/U6O0MVvkm3
Your AI vendor's vendor's cloud provider got breached.
Fireworks, Oct 24: dev machines hit via a third-party cloud host. Notion, downstream: under 1% of workspaces. Their advice: rotate any credential you pasted into Notion AI.
Where else did you paste a key?
The three checks it maps to: https://t.co/NK5BMqfmL2
Ship Check Kit has your coding agent run 44 launch and security checks, RLS policies included, with a fix for every failure. $29. https://t.co/U6O0MVvkm3
One missing check in an invitations RLS policy.
Dosu, Nov 2: a free account invited itself as owner of their internal org, repointed integrations to grab creds, then a PR got CI to hand over prod cloud keys. Under two hours.
Tested your invite insert as a second user?
What to rotate, platform by platform (15 named so far): https://t.co/3BF9VZ0MHR
Ship Check Kit has your coding agent run 44 launch and security checks, leaked keys included, with a fix for every failure. $29. https://t.co/U6O0MVvkm3
The breach email isn't the incident. The unrotated key is.
Nov 8, Tensorlake said a host breach "a couple of weeks ago" exposed their env vars. They hadn't rotated them. Someone used them to push branches and run GitHub Actions to probe secrets.
Got that email? Rotated yet?