JI-VENTURES

Me and my AI agent team, building a factory of products in public. ShipMender (security scans for vibe-coded apps) is the first of many.

[JI-VENTURES @itsjasonissac]

@suraj_sharma14 Honestly somewhere around stage 9. I'd sneak in a stage 3.5 though: open your live app logged out and try reading a table with just the public key. Two minutes, and it's the cheapest boss fight you'll get.

[JI-VENTURES @itsjasonissac]

@ArianHeydarix Good shout, a policy that's "on" but allows everyone is basically an unlocked door with a lock fitted. Testing as a logged-out visitor is the only way to be sure.

[JI-VENTURES @itsjasonissac]

1/ Someone on Reddit asked if an API key in the frontend is fine because their app only has 3 users.

Fair question. I've wondered the same. But the bots scanning public JavaScript don't check your user count first.

5 checks I run before anything ships:

[JI-VENTURES @itsjasonissac]

Railway's own notice: https://t.co/p7DyKbF5Cr

[JI-VENTURES @itsjasonissac]

Railway (the app hosting platform) says someone may have got into customer environment variables, the secret settings where API keys and database passwords live.

If your app runs on Railway, here's what I'd do today:

1. Make new API keys (OpenAI, Stripe, the lot) and delete the old ones
2. Change your database password
3. Check your AI and payment dashboards for usage you don't recognise
4. Put spending limits on your AI keys

Boring jobs, but cheaper than a surprise bill.

[JI-VENTURES @itsjasonissac]

https://t.co/OSdMbe0XbE

[JI-VENTURES @itsjasonissac]

My AI team doesn't send anything until I've read the exact words.

Saw a post this week from someone whose assistant sent a business email without showing them a draft. Trust gone in one click.

We work on the building, ops and posts together. They draft, research and flag. I read it, change it, and press send.

Anthropic's own report says 80% of agent session time is spent waiting on a human. I'm fine being that human. Slower, but nothing goes out I wouldn't say myself.

Where do you draw the line on letting agents act on their own?

[JI-VENTURES @itsjasonissac]

6/ Set spend caps and billing alerts, and turn on secret scanning in GitHub. If a key does leak, rotate it first and investigate second.

That's it. No shame in asking. I'd rather people ask than find out from the invoice.

[JI-VENTURES @itsjasonissac]

5/ The anon key is only safe if row-level security is on. Otherwise it's a front door with no lock.

[JI-VENTURES @itsjasonissac]

@isjackback That's brilliant. Did you do anything on purpose to get picked up by ChatGPT, or did it just start happening?

[JI-VENTURES @itsjasonissac]

@sophie_launch Saying no to the tidy up is the only way to bed before sunrise.

[JI-VENTURES @itsjasonissac]

@snskritinaruka Binning slop tests makes sense. I'd still keep a small hand checked set around the bits that hurt quietly when they break: auth, who can read which rows, payments. Those are the tests an agent should never be allowed to rewrite just to go green.

[JI-VENTURES @itsjasonissac]

@theo Remote first always ends up like Minecraft redstone. You just wanted a door that opens, and now there's a wall of repeaters and a comparator nobody remembers placing. Worth it once it works though.

[JI-VENTURES @itsjasonissac]

@AnatoliKopadze The reply guy bit is the part everyone will scroll past. It's the slowest channel there is, and that's why it works early. Nobody can copy a hundred real conversations.

[JI-VENTURES @itsjasonissac]

@GergelyOrosz Running a small business with a few AI agents, the model is the last thing I worry about. What matters is the plumbing: what each agent is allowed to touch, and who signs off before anything goes out.

[JI-VENTURES @itsjasonissac]

@supabase Light mode and tables without RLS. Both technically supported, both make me squint.