What garak also does
This course taught the scanning loop and the parts of it an application needs. Garak is larger, and knowing the shape of the rest is worth ten minutes.
| Not covered here | What it is |
|---|---|
| About 190 probes | This course ran a handful. The rest are the catalogue you are paying for |
| About 117 detectors | Including several that run their own models |
| The report HTML | A scored summary written beside the JSON, grouped by risk |
| Tiers and the z-score | Garak's own scoring of how much a result matters |
| The intent service | The risk categories in the banner, and how results roll up to them |
| Harnesses | probewise and pxd, which decide how probes and detectors are paired |
| Around 44 generators | Every model API, plus rest, and a browser |
| Payloads | Reusable content that several probes draw from |
| --parallel_attempts | Concurrency, which the run suggests when your target supports it |
Where it sits beside the other tools
PyRIT is the other red teaming course here. Garak is a scanner: a catalogue of known weaknesses, run in bulk, reported as rates. PyRIT is a harness: you build the attack, control the conversation, and keep everything in a database. Garak answers does my system have any of these known problems; PyRIT answers what happens if I try this.
Guardrails AI and NeMo Guardrails are the fix rather than the finding. A scan tells you an answer got through; a guardrail is what stops it next time. Running a scan against a system with rails, and again with them off, is how you show a rail is worth its latency.
- Open the HTML report from any lesson and find the risk grouping.
- Read the probe list and pick five you would run against your own system.
- Take one hit from a scan and write the rail that would stop it.
Little by little, you're building something great.