CollieAi

About CollieAi

CollieAi is a security proxy that sits in front of an LLM application and filters both the prompts going in and the responses coming out. Adoption is deliberately trivial: you change the base_url on your existing OpenAI-compatible or Anthropic-native client and keep every other line of code, so there is no SDK to adopt and no architectural change. Behind that one line it runs a three-layer detection pipeline — deterministic pattern matching, machine-learning classifiers, and a generative model that reasons about intent — across nine rule types. The threats it names are the practical ones: direct prompt injection and the indirect kind that arrives inside a RAG document or a tool result, jailbreak and safety-bypass attempts, PII and financial-data leakage, exposed secrets matched by shape (sk-, ghp-, AKIA, JWT), malicious or obfuscated URLs, base64-hidden payloads, Unicode evasion via homoglyphs and zero-width characters, and unsafe or off-policy model output. It runs in three shapes — drop-in proxy, real-time SSE streaming, and async jobs — and in either Monitor mode, which logs without blocking, or Enforce mode. Operationally it provides a real-time dashboard with logs, analytics and threshold alerts, SIEM export and a full audit trail. It can be run as cloud SaaS, self-hosted or on-premise, with an open-source core.

Key Features

Tags