FuzzyAI

Why we included this project

Anyone shipping an application on top of an LLM API eventually wonders how easy it is to talk the model out of its safety rules. FuzzyAI answers that question with automated fuzzing: it throws a set of jailbreak-style attack patterns at a model endpoint and reports which ones get through, so weak spots show up before someone else finds them. Security researchers get the bundled attacks for testing hosted or locally served models, and developers can fold the fuzzer into their own pipeline to check the third-party models their app depends on. Setup stays light, since you can install it via pip, run it against a local ollama model, or use the experimental web interface. For anyone accountable for LLM security, that turns a vague worry about jailbreaks into a repeatable test run you can document and act on.