A scan of 159,041 Lovable projects found 200,000+ private chat messages and 1,000+ passport records readable without a login.
The public Supabase key isn't the problem. Missing row-level security is.
A 20-minute check for your own app:
2
1
A feedback doc is ordered for writing. A tracker is ordered for fixing.
One review became 117 numbered rows: a screenshot, a batch and a status on each, plus the rules for every agent session at the bottom.
How to set up yours:
1
Hightouch tested decision models as LLM judges: up to 90x cheaper and 10x faster than a frontier model, about 6 points less accurate on hard cases.
The trick: send the low-confidence calls to Opus. Frontier accuracy at over 4x lower cost:
1
Rules in AGENTS.md are suggestions. The agent reads them, then guesses what "run the tests" means.
A script with a pass/fail result is a constraint.
The case for swapping your agent's markdown for ./scripts/lint.sh and make test:
2
2
