Securing Agentic AI Developer Day: Evaluating Guardrailed Systems and Models with `garak`

Why it matters

Guardrails reduce risk only if the same adversarial probes succeed less often against them. This notebook reruns the workshop’s garak_demo probe set against the guarded model so the two runs can be compared.

What to practise

Read the garak config: the probe_spec that lists the probe families (Markdown image exfiltration, GCG suffix, DAN and latent injection), generations, and the model_type: guardrails generator whose model_name is a Guardrails config directory rather than a NIM model. Run garak --config and compare the HTML report with the unguarded run. Then, in the optional step, check the hit log for probes that got through and edit config.yml or flows.co to stop the sample jailbreak.

How it fits

This is 9.4.2’s matched comparison with garak, with the same probes and only the guardrails changed. A probe report shows attack coverage, not legitimate task completion. Measure that separately.