Promptfoo and Ollama: red teaming setup and reports

This beginner tutorial walks through Promptfoo's red teaming interface with an Ollama model as the target. The presenter installs Promptfoo with Homebrew, loads an example configuration, and explains how target settings differ for models and HTTP endpoints. The demonstration uses an abliterated model, which the presenter describes as having safety guardrails removed.

The configuration walkthrough covers attacker personas, application context, plugin presets, custom intents and policies. It also explains static and dynamic strategies, including single-turn Meta Agent and multi-turn Hydra attacks. Viewers see concurrency and delay controls, language selection, and YAML export. In the demonstrated interface, the YAML preview does not allow editing; the presenter saves the file for changes in an external editor.

The report walkthrough distinguishes attack success from successful defense and examines prompts, responses, risk categories and CSV exports. Although the target is a local LLM, the presenter notices OpenAI calls for attack generation and explains where to configure another attack model. This run therefore does not establish an offline workflow.

Several settings remain unresolved, including custom variables and parts of auto-discovery. The scan finishes in 2 hours 35 minutes against a 7-minute estimate, with 55 passes, eight failures and a reported 12.7% attack success rate. These are results from this demonstration, not general benchmarks.