Koreshield

Every support workflow starts with text from someone outside your company. The input scan inspects that text before it enters the model context. It looks for instruction overrides, attempts to pull private data, and other hostile patterns, then returns a decision with a severity and a confidence score. This is the first place teams add prompt injection protection, and it is often the easiest to verify with real tickets.

Support agents rarely answer from the model alone. They pull policy pages, order history, ticket notes, and CRM snippets into the prompt. The RAG scan checks selected documents and chunks before prompt assembly. That step matters when a knowledge base can be edited by customers, partners, or a compromised internal account. A poisoned article that never reaches the prompt cannot steer the answer.

The third checkpoint evaluates proposed tool calls. A refund, a payment link, an account closure, or a data export is a business action, and the model does not know your approval limits. Koreshield reviews the arguments and the authorization context before execution, then returns a decision your application can follow. Your application still owns identity, permissions, and the final call.

Detect mode records threats without interrupting traffic. Each event includes a request ID, a severity, a confidence score, and the indicators that fired. Support and security teams can review false positives and misses against live behavior before changing what customers see. The approach fits companies that need proof before they put a block in front of a revenue workflow.

Enforce mode stops requests that violate policy. It is designed to be turned on after detect mode has shown what normal traffic looks like. The site tells teams to define the customer message, the operator escalation path, and the retry behavior first. That advice is practical: a blocked refund with no fallback creates a support ticket, not a security win.

The hosted console handles keys, evidence review, policy changes, mode switching, and validation. Scan keys are separate and revocable, and the scan-only OpenAPI contract lets an agent platform import the three checks without exposing billing, team administration, or evidence review as model-selectable tools. Koreshield can run as a hosted service or as a FastAPI service with PostgreSQL inside your own infrastructure.