Skip to main content

Pandamonium (experimental)

The Pandamonium (Prompt-based Automation for Navigating Discovery of Attacks, Misuse, Opportunistic Nefarious Intents, and Uncovering Model Exploits) strategy is an advanced automated red teaming technique that dynamically generates single or multi-turn conversations aimed at bypassing a target model's safety measures.

warning

This is an experimental strategy currently in development by the Promptfoo team.

warning

This strategy does not have a token limit and will continue to run until it finds a jailbreak or is stopped.

Implementation​

Use it like so in your promptfooconfig.yaml:

promptfooconfig.yaml
strategies:
  - id: pandamonium

How It Works​

Pandamonium uses a specialized attack agent that:

  1. Analyzes the target model's behavior and responses
  2. Dynamically generates and refines attack approaches
  3. Pursues multiple exploitation pathways in parallel
  4. Adapts tactics based on the model's safety mechanisms

Unlike other strategies, Pandamonium will continue running until it either successfully jailbreaks the model or is manually stopped.

For a comprehensive overview of LLM vulnerabilities and red teaming strategies, visit our Types of LLM Vulnerabilities page.