Мутации промптов для adversarial-тестирования

★ 7.7 · testing

attack-mutator is a Claude Code skill that generates variations of test cases and prompts for adversarial testing using semantic, encoding, and visual mutation techniques. Supported transformations include paraphrase, synonym_swap, formality_shift, base64_wrap, ROT13, unicode homoglyphs, zero-width character injection, and leetspeak. Each mutation is scored by an algorithm that weighs semantic preservation, surface novelty, and length ratio, then returns a ranked list of candidates. The skill is invoked via /mutate with optional --types and --best-of-n flags, and works with any model configured in Claude settings. It is designed for security and QA teams who need to stress-test filter bypass resistance, probe prompt injection vulnerabilities, and maximize edge-case coverage.