Argon Guard
Argon Guard is a trust and safety agent for AI builder platforms, the products that let anyone ship a site, a game, or an automation in seconds.
Core trick
Runs rules and a cheap model on every deploy, escalating only uncertain or rule-flagged pages to a high-thinking model for the final verdict.
Tags
Description
Argon Guard is a trust and safety agent for AI builder platforms, the products that let anyone ship a site, a game, or an automation in seconds. That speed also lets people ship phishing pages, scams, and hate sites, and the platform hosting them takes the blame and the legal risk.
Every deploy gets scanned before it goes live. A rules pass and Gemini 3.8 Flash (low thinking) score each page fast and cheap. When Flash is unsure or the rules fire, the scan escalates to Gemini 4 Argon at high thinking, which makes the final call and explains why. It covers phishing, scams, hate and harassment, and violent extremism, and it is built to leave education, news, and counter-speech alone. It also looks outward: an impersonation finder generates lookalike domains, checks certificate transparency logs, and compares live pages against your site to catch copycats.
On a 250-page sample of the public PhreshPhish dataset, the tiered setup scored 100% precision with 0 false positives, and only 36% of scans needed Argon. On our hate and extremism set (20 pages, half of them tricky benign pages), it scored 100% precision and 100% recall with 0% false positives, versus 90/90/10% for rules alone.
It works today as a Vercel integration, with any-URL scanning for other platforms. Live app: https://trust-safety-agent.vercel.app
Pitch deck: https://canva.link/w1gx56h9995yzmp
DasilvaKareem/trust-safety-agent★ 00 forksTypeScriptMITCreated Oct 25, 2026 · last push Oct 25, 2026
README excerpt
# Argon Guard [](https://github.com/DasilvaKareem/trust-safety-agent/actions/workflows/ci.yml) A hackathon website-safety assistant built with Next.js, Neon Auth/Postgres, Google Gemini, and [eve](https://www.npmjs.com/package/eve). Connect a website or selected Vercel projects, scan public pages, and review the evidence behind potential phishing, impersonation, scams, hate, or extremist content. **[Open Argon Guard](https://argon-guard.vercel.app)** · **[Setup and deployment](DEPLOYMENT.md)** · **[Architecture](docs/ARCHITECTURE.md)** · **[Evaluation results](bench/results.md)**  Built for the CV Gemini 4 Argon Hackathon. This is a working prototype, not a security certification. Scans are read-only; the app does not block or take down deployments. ## What works - Agent-led onboarding with a mouse-tracking avatar, Google sign-in, and private workspaces. - Public-URL scans, with ownership verification before enabling monitoring. - Vercel OAuth connection with read-only project/deployment access and signed deployment webhooks. - Rules plus a fast Gemini pass, escalating to a deeper model when warranted. Missing/failed AI analysis falls back to rules and is identified in the results. - Scan history, review decisions, and content groups using SHA-256 and SimHash. - An outward impersonation search using domain vari