Adam Persson

Built this because I kept watching developers ship agentic AI features with no security layer. The tools that existed were either paywalled, hard to self-host, or didn't do multilingual detection well. Defend is local-first by design. The classifier runs on Qwen 0.5B, fine-tuned on ~70k real + synthetic samples, covers 30+ languages. If you need higher confidence, you can route any module...