roadmap
- 01
- 02
- 03
- 04
- 05
- 06
- 07
- 08
Scan the models and LLM apps your team ships with garak before release, and report what it finds without overclaiming. Covers probes, detectors, targets, run sizing, scoring, triage and CI gates.
154 of 154 questions in your plan: 0 known, 0 in progress, 0 due, 0 skipped, 154 new. 0 cut, outside your plan.
garak usually comes up in loops for AI security, AI red-team and LLM platform roles, and for application engineers who own a chat feature and must show how they would test it before release. It tends to arrive inside a technical round on evaluating an LLM system, or as a scan report put on the table with the question "would you ship this?"
Interviewers probe what a garak number does and does not attest to. A probe whose prompts come from a fixed public list may pass because the model was tuned on that list. A detector that matches refusal strings can rule a successful attack a pass, or flag a harmless reply as a hit. One generation per prompt turns an intermittent failure into a coin flip. A generator pointed at the bare model endpoint never sees the system prompt, the session history or the JSON envelope your real app adds.
A junior answer explains the plugin types and can run a scan. A senior answer starts from a report and works backwards: which probes ran and which did not, what the detector's error rate does to the failure rate, whether an empty uncertainty interval means the sample was too small to quote, and which of these results is sound enough to block a release.
The plan starts with the defender's map of what can go wrong in an LLM application, the risk categories a scan is meant to test. It then turns to the attacks garak's probes encode: prompt injection, the jailbreak families and why a published one loses its effect over time, and input built to slip past a screen, which is what a buff does to a probe's prompts. They appear here as the failures your own system is tested for.
Two measurement ideas come before the tool, because garak's results are read through them. A detector is a classifier, so precision and recall describe its errors. A fixed prompt list behaves like a public benchmark, so contamination and saturation limit what a pass on it proves.
garak itself starts with its plugins: what probes send, how detectors rule each reply, how a generator reaches a deployed app, and what you take on when you write your own pair. Sizing the run comes next. It opens on why an attack that works only sometimes still counts and how sampling error distorts a rate, then turns to garak's own levers: which probes to run on a budget and how many times to repeat each prompt.
Reading the output follows. Safety evaluation supplies the frame for an attack-success rate, then garak's absolute and relative scores and the work of triaging hits by hand. The plan ends where a scan becomes routine: a regression suite in CI with pinned settings, a baseline and a cost budget per run.
If you build the LLM feature and have been asked to add a scan before it ships, start at the plugin model and the generator questions, because wiring garak to your app's real endpoint is where first scans go wrong, then work through sizing and triage. Coming from application security, take the attack classes first; the scanner's vocabulary assumes them. For a senior or principal interview, narrow level to that range and spend your time on the measurement, scoring and gating stretches, where the questions are about what a number may be used for. The plan covers one scanner run against systems you own or are authorised to test. Multi-turn attack orchestration, manual red-team technique beyond the classes listed here, and writing up a full engagement belong to the [AI Red Teaming roadmap](/roadmaps/ai-red-teaming). Building the defenses a scan measures, such as guardrails and injection containment, is part of the [AI Engineer roadmap](/roadmaps/ai-engineer).
If you build the LLM feature and have been asked to add a scan before it ships, start at the plugin model and the generator questions, because wiring garak to your app's real endpoint is where first scans go wrong, then work through sizing and triage. Coming from application security, take the attack classes first; the scanner's vocabulary assumes them. For a senior or principal interview, narrow level to that range and spend your time on the measurement, scoring and gating stretches, where the questions are about what a number may be used for.
The plan covers one scanner run against systems you own or are authorised to test. Multi-turn attack orchestration, manual red-team technique beyond the classes listed here, and writing up a full engagement belong to the AI Red Teaming roadmap. Building the defenses a scan measures, such as guardrails and injection containment, is part of the AI Engineer roadmap.