Sunglasses is an open source filter that blocks prompt injection, MCP tool poisoning and credential exfiltration before they reach your AI agent. 1,124 tested patterns across 84 attack categories, shipping daily. Runs fully local. No cloud, no account, no telemetry. pip install sunglasses, gate PRs with the GitHub Action or paste any GitHub repo at sunglasses.dev/scan to see what your agent would swallow. MIT licensed with a reproducible benchmark inside the repo.

Hey Product Hunt 👋

I'm Azad. I come from a business background, not engineering. I started building Sunglasses in February because I was running AI agents on my own machine and realized nothing checks what they read.

Agents treat everything as instructions. Files, emails, web pages, tool outputs. Attackers figured this out and write for the AI now, not for you. Sunglasses sits in front of your agent and scans untrusted input before the agent acts on it.

What makes it different.

Runs 100% local. No cloud, no account, no telemetry. MIT licensed.

1,124 tested attack patterns across 84 categories as of this morning. We ship new ones daily.

Honest about limits. We publish what it does NOT catch at sunglasses.dev/what-sunglasses-catches-vs-does-not-catch. A security tool that hides its gaps is worse than no tool.

The benchmark ships inside the repo and reproduces with one command. 78.7% precision and 97.4% recall on the labeled set.

Try it in 10 seconds. Paste any GitHub repo at sunglasses.dev/scan and see what an AI coding agent would swallow. Or pip install sunglasses.

If you break it, tell me. Bypasses become detection patterns with credit to you in the changelog. Ask me anything.

the daily pattern updates are genuinely sharp, treating it like a living threat feed rather than a one-off library. love that the benchmark lives in the same repo so i can actually trust the numbers.