Lovable got caught exposing prompts, chat history, and even credentials from public projects through its.
here’s the screenshot that shows how far the exposure went.
okay so “public project” shouldn’t mean “public prompt transcript + creds,” that’s a pretty brutal failure mode for a tool that literally encourages people to paste half their app into a chat box. it’s the live-mic problem: once it’s in the signal chain, assume it can leak somewhere you didn’t intend.
i’m curious if Lovable had any secret-scanning or redaction on the prompt/chat stream at all, or if the safety model was basically “pls don’t paste keys.” because even careful people slip once, and then it’s cached in history forever.
i’ve been doing the same thing you mentioned: aggressive placeholders in anything that might end up in a prompt, plus keeping real values strictly in local .env / password manager. even a fake-but-realistic token like sk_live_REDACTED is enough to keep your brain from autopiloting into pasting the actual one.
“public project” meaning “public prompt transcript” is such a nasty footgun here, because chat feels disposable but it’s basically a forever-log.
The “live-mic” thing is exactly it — once it’s in the prompt stream you’re trusting a whole conveyor belt of UI, logs, exports, and “helpful” share views to never cough it back up. I genuinely don’t know if Lovable is doing any secret-scanning/redaction, but relying on “pls don’t paste keys” for a product that encourages copy/pasting half an app is… optimistic.
The fake-but-realistic token trick has saved me a couple times too. If it looks like sk_live_REDACTED my brain hits the brakes, vs 11pm autopilot pasting the real one.
