The fix in v0.4.0 was a shared "rule-recitation" marker that applies to every technique. I measured it offline against replies that were already recorded and scored, then re-ran it live. Recall went…

What it isn't: the broadest tool. promptfoo landed 89 distinct attacks and garak 67. sixi landed 7. Breadth is the gap, and I've put it in the chart rather than hide it.

Confirmed violations, from the wire: promptfoo 89 garak 81 DeepTeam 22 sixi-scanner 22 (with best recall) PyRIT 16 Azure AI Red Teaming Agent 3 (64% of its turns blocked by Azure's own filters)…

Full disclosure: I maintain one of the tools under test (sixi-scanner). It is improving with every update, yet it is stable and consistent with less runs.