Make any LLM find vulnerabilties & bugs

Works on 120b, 20b and even 8b models

4 followers

Works on 120b, 20b and even 8b models

4 followers

A small model alone is a weak bug finder: it hallucinates issues and misses the real ones. RedMirror Reflection turns it into a real one. It guides the model to where bugs hide, then searches every reachable state to discover the exact sequence that breaks it, returning GROUNDED with the attack path or a proof it’s safe. The search costs no tokens, so a 20B model matched a 120B at 96% recall on the 440-case OpenSSF benchmark. One line: redmirror-reflect init First month free, then $10/seat.

This is the 2nd launch from Make any LLM find vulnerabilties & bugs. View more

RedMirror: make any LLM find real bugs

Launching today

Make any LLM find real bugs in your code, and prove them

Give any model the power to find real bugs and prove them: it guides your agent to where bugs live, then a compiled kernel searches every reachable state and returns a replayable attack path, or a proof it's safe. New: it now also finds injection bugs and lands the exploit on a live app. On the 440-case OpenSSF benchmark a 20b model matches the 120b at 96%. Runs offline, your model. Free 7-day trial, then $4.99/mo.

RedMirror: make any LLM find real bugs gallery image

Launch tags:Developer Tools•Artificial Intelligence•Security

Launch Team

Wispr Flow: Dictation That Works EverywhereStop typing. Start speaking. 4x faster.

Promoted

Maker

📌

Hey Product Hunt 👋 I wanted to point a model at my code and have it actually find real vulnerabilities. On its own it can't: it invents bugs that aren't there and walks past the ones that are. RedMirror Reflection is what makes it a real bug finder. Since we were last here we've shipped a lot: it now also finds injection bugs (SQLi, command injection, path traversal, SSRF, XSS) and can guide your model to land the exploit on a live app, and there's a free 7-day trial now, no card. Here's how it discovers bugs instead of just checking them. It guides the model to where bugs actually live (the vulnerable relations and the angles that matter for that subsystem: authorization, conservation of value, ownership, path handling, idempotency), then a compiled kernel searches every reachable state and surfaces the exact sequence that breaks it. You get back GROUNDED with the concrete attack path, or a proof the code is safe. Evidence, not an opinion, with a stamp you can replay. The part that still surprises me: because that exhaustive search costs zero tokens, the model barely has to be smart. On the 440-case OpenSSF benchmark, gpt-oss-20b found 96% of the real bugs, the same recall as the 120B. A local qwen model ran the whole thing on one GPU. It drops into Claude Code, Cursor, Codex, Gemini CLI and pi with one line: redmirror-reflect init . The binary is an MCP server, so your agent drives it natively. Bring your own model, cloud or fully offline, your code never leaves your machine. Free 7-day trial, then $4.99 per seat. I'd love your feedback, especially which small model you'd point at it and what you'd want it to find next. I'm here all day.

Report

13h ago

Previous Make any LLM find vulnerabilties & bugs Launches

Make any LLM find vulnerabilties & bugsWorks on 120b, 20b and even 8b models

Launched on August 10th, 2026