Show HN: Built firewall for LLMs after prompt injection bypass GPT-4s guardrails(github.com)

1 pointsby Dushmesh_DS5 months ago1 comments

Show HN: Built firewall for LLMs after prompt injection bypass GPT-4s guardrails

https://github.com/WardLink/TrustLayer--Security-Control-Plane-For-LLM-AI

After watching prompt injection attacks bypass every major LLM's safety systems, I built TrustLayer.

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.

1 comments

SamiBuilds5 months ago

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.