
What happened
A non-engineer built an AI receptionist app for the 5th Agentic AI Hackathon with Google Cloud by vibe coding, then made Gemini act as a security advisor to design 20 safeguards, from blocking emergency inputs before the LLM to hiding API keys behind a relay server.
Why it matters
The 20 safeguards are meant to keep the app safe for visitors to medical and corporate sites, covering risks like prompt injection and personal data leaks—an approach other non-engineers building with AI might learn from, though the body presents it as one person's experience, not a proven standard.
Summaries like this, in your inbox every morning.
The author is upfront about not being an engineer, and that perspective shapes the whole account. The project was built for the 5th Agentic AI Hackathon with Google Cloud, where the app—an AI receptionist avatar that greets and sees off visitors on medical and corporate websites—was created entirely through vibe coding. The turning point came when 32 3D avatars vanished from the screen because an external service stopped, leaving the reception area empty. That scare, along with smaller incidents like a key written into the README and Cyrillic characters mixing into a key, pushed the author to treat security as something to build into the system rather than remember to do. Working with Gemini as a security advisor, the author ended up with 20 safeguards split between system design (#1–#16) and lessons from real mishaps (#17–#20). The through-line is a shift from relying on the AI's own judgment to placing deterministic code around it: blocking emergency inputs before they reach the LLM, forcing source citations to come from a registered list, and distinguishing permission errors from 'no data' so users are not misled. For a non-engineer, the final check is simpler—open the app on the same device and in the same way a user would, and look with your own eyes.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Docker published an analysis of major AI coding agent incidents from October 2024 to February 2026, including…

Using ClickCapture screenshots of Robocopy Scheduler, Gemini 2.0 Flash-Lite produced a structured parameter ta…

After two ✓-marked items turned out wrong post-publication — a nonexistent "創業枠(200万円)" and a 3-month photo ru…

A weekend solo engineer asked Claude Code to move a skill that triggers CodeRabbit reviews every 65–70 minutes…

Drawing on 11 years as a 臨床工学技士, the author proposes six fields for AI agent approval logs: when, who, what ta…

A solo Amazon reseller uses Claude's scheduled tasks to run shipping requests, shipment notices, inventory syn…
