A real laboratory on your computer. Challenge a model, compare its defenses, and turn every result into something you understand.
On your device · No API key required · Five languages
Start with the lesson, follow a concrete example, then test what changes in the laboratory. Learn to distinguish an injection attempt, model behavior and an actual security impact.
OSAFISPDF + PPTXPrompt Injection
Mechanisms, examples, defenses and evidence-based evaluation in one course.
Choose a language and open the courseSelect the PDF or PowerPoint language separately from the site language.
From direct prompt injection to indirect inputs and tools, one experiment becomes the starting point for your next question.
Compare a prompt across different architectures and observe what changes when you move the boundary of trust.
Adjust difficulty, inspect the execution trace and follow your progress in a report saved on your own machine.
Run with Ollama on your computer. Once dependencies and the model are downloaded, experiments can work offline.
Choose your operating system. We will take you from the first download to a running model.
You can explore this page on mobile, but a computer is required to run the laboratory.
Run this beside the ZIP and compare the output with the hash below. A checksum is not a publisher signature.
da9f6638952413acccd39425e1e27422f1be1501869119208d009d3a87468a0dPractical starting estimate; 16 GB for more headroom
Initial budget for tools and the approximately 1.9 GB model
Preferably 4 cores; a compatible GPU is optional for speed
RAM, CPU and storage figures are practical estimates, not measured or guaranteed minimums. Context length, model, OS and other running apps affect resource use.
This laboratory runs with a local model, whose responses can vary or be incorrect. If you encounter a problem, have a suggestion or find a bug, email us to help improve future releases.
report@osafis.orgInclude your OS, Python version, model name and error message. Remove personal or sensitive information before sending.