My prompt-injection firewall showed a perfect 9/9 score. It was fake.
Project Details
Learning project, would love feedback. Stack: regex rules -> Gemini classifier (JSON + Pydantic + retries, fail-closed) -> output check (canary, secret patterns, 8-word system-prompt match) Data: 25 cases from deepset/prompt-injections + 12 output cases I wrote with AI help. Results - Rules only: 2/9 attacks caught - Rules + Gemini: 8/9 dev, 0/6 false positives - Hidden: 3/3 caught, 0/7 false positives (only 3 attacks, so weak evidence) Things that went wrong 💀 - Wrong model name ->
Similar Projects
I built a job posting alert tool because I could never tell if a job posting was actually real and w
Real, current tech job postings straight from each company's own careers page. Set an aler...
I built ZyMetrix because I was tired of using 10 different sites for basic tools
Convert files, edit PDFs, scan barcodes, process images, share text and use developer tool...
Play how many words u can in 90Sec. Have u played word Chaining of last letter game ever? I was use
Word Chain — built on Replit. Update this description to reflect the app.
I built a free online dev tools website because I was tired of every tool asking me to sign up
Free developer utilities online. No sign-up, no tracking on your data. JSON, JWT, regex, t...
my site took 3 minutes to load. that was the whole problem
Every number comes with its confidence band, and when the data is not enough the page says...
I built a map for every food place I save on TikTok and Reels, because the saved folder was useless
Download BiteVault for iPhone or Android, or use the web app in any browser. Save recipes...
