wtf is ai-firewall?

AI Firewall is a security gateway that sits between your applications and large language model APIs like those from OpenAI, Anthropic, and Google. Its job is to inspect outgoing prompts before they leave your network and incoming model responses before they reach your users. It redacts sensitive data such as AWS keys, API tokens, credit card numbers, and social security numbers. It also filters out prompt injection and jailbreak attempts, and it tags every request with a risk category under the EU AI Act, writing those tags to a persistent audit log. The project has two independent pieces. A Go edge proxy handles the fast, low latency work: authenticating requests, scanning for attack phrases using a string matching algorithm called Aho-Corasick, redacting secrets with regex patterns, and intercepting streaming responses to scrub secrets even when they are split across chunk boundaries. A TypeScript management plane handles multi-provider forwarding, audit persistence, API key management, and a React operations dashboard with a playground, rules engine, stream visualizer, and live benchmark screen. The README is unusually honest about what works and what does not. The TypeScript gateway, Go proxy, and dashboard are all marked as working and unit-tested. A live benchmark screen measures precision, recall, and latency at runtime against a 30-case labeled corpus, with no pre-recorded numbers. However, an ML classifier for detecting paraphrased or multilingual attacks is not implemented yet, and distributed deployment, rate limiting, and PostgreSQL storage remain on the roadmap. The current detection layer is purely deterministic, relying on keyword matching and regular expressions. This makes it fast and precise on known patterns but weak against paraphrased or multilingual attacks. The benchmark screen exposes this gap on purpose, and adding an ONNX transformer classifier is the top roadmap priority. The code is available under the Business Source License 1.1, which permits free non-production use and production use up to 10,000 inspected requests per day. Offering it as a competing hosted service requires a commercial license. The license automatically converts to Apache 2.0 on July 23, 2030.