Show HN: Built firewall for LLMs after prompt injection bypass GPT-4s guardrails(github.com)
github.com
Show HN: Built firewall for LLMs after prompt injection bypass GPT-4s guardrails
https://github.com/WardLink/TrustLayer--Security-Control-Plane-For-LLM-AI
1 comments
Interesting approach! In our work on API security, we've been exploring tools that analyze OpenAPI specs for semantic intent, catching risky logic before deployment. The idea of a 'guard layer' resonates—especially for preventing edge-case exposures. Curious how you handle multi-step prompt manipulations in production scenarios.
It's a simple API that: 1. Scans prompts for injection attacks in real-time (heuristic + OpenAI moderation) 2. Detects when agents drift from their intended behavior 3. Has a kill switch for production incidents
Built for production use. Free tier available. Open source docs and examples.
The problem: Prompt injection is still breaking LLM apps in production. Most guardrails can be bypassed with simple tricks.
The solution: API-first security layer that sits between your app and the LLM.