Koreshield

Koreshield is an AI security tool by Koreshield Labs (UK) that inspects untrusted content flowing through AI customer support agents before it becomes a trusted instruction. Your server sends each message, document or proposed action to it and gets back a decision with a severity level and the evidence behind it, while you keep your current model provider.

What is Koreshield used for?

It closes the gap between model safety and the business decision around it. A model vendor guards the model, but it can't know who the customer is or what your approval limits are. Koreshield watches three trust boundaries: customer input, retrieved context such as tickets, CRM notes or RAG documents that may hide indirect prompt injection, and proposed tool calls with their arguments and authorization context. It also scans model output and redacts secrets and personal data.

How does Koreshield work?

It isn't a proxy. Your backend calls the API, directly or through the Python and TypeScript SDKs, right before each sensitive step. You begin in detect mode, which logs threats without blocking traffic, then validate with benign and adversarial cases and switch on enforce mode once fallback behavior is defined. Every decision keeps indicators, confidence and timing for review. The service runs on FastAPI with PostgreSQL and connects from the hosted console or your own stack, with a ready-made Melaya integration.

Who is Koreshield for?

• E-commerce and DTC teams whose agents can change orders or issue refunds.
• B2B SaaS companies with assistants that read private workspaces.
• Financial services firms handling sensitive data in AI support.

Koreshield pros and cons

On the plus side, there's no model swap, rollout is gradual and every decision leaves an audit trail. On the downside, the company says it doesn't replace your authorization layer, doesn't promise full attack coverage, can't inspect images or arbitrary files yet and isn't meant for static FAQ bots. The site and docs are in English.