promptfoo-strategies-dynamiclisted

# Dynamic Red Team Strategies Dynamic strategies use an attacker LLM to mutate adversarial inputs through iterative refinement. Multiple calls to both attacker and target models determine the most effective attack vector. Higher success rates (40-90% ASR increase) than static strategies, but more resource-intensive. Stop after exhausting token budget or on successful harmful output. ## Use this skill when - Iteratively refining prompts to bypass security controls (`jailbreak`) - Building custom attack taxonomies with persistent memory (`jailbreak:meta`) - Chaining multiple jailbreak techniques from research papers (`jailbreak:composite`) - Running Tree of Attacks branching exploration (`jailbreak:tree`) - Using Best-of-N parallel sampling from Anthropic research (`best-of-n`) - Testing academic authority bias framing (`citation`, `likert`) - Testing mathematical notation-based attacks (`math-prompt`) - Testing structured markup authority exploitation (`authoritative-markup-injection`) - Running gradient-based GCG adversarial prompt search (`gcg`) ## Do not use this skill when - You need deterministic, low-cost encoding bypasses → use `promptfoo-strategies-static` - You need multi-turn conversation attacks → use `promptfoo-strategies-multi-turn` - You need indirect prompt injection via web pages → use `promptfoo-strategies-indirect-injection` - You need custom or regression strategies → use `promptfoo-strategies-custom-regression` - Cloud access is unavailable AND the strat