Whimsical Strategies Skill — AI skill for Claude Code

whimsical-strategies-skill

A [Claude Code agent skill](https://code.claude.com/docs/en/skills) that generates **out-of-distribution adversarial prompts** for AI agents, using the methodology from Microsoft Research's *"Whimsical strategies break AI agents"* ([May 2026](https://www.microsoft.com/en-us/research/articles/whimsical-strategies-break-ai-agents-generating-out-of-distribution-adversarial-strategies-at-scale/)).

What it does

Seeds adversarial-tactic generation with **random external knowledge** (Wikipedia by default) so the LLM surfaces cross-domain analogies a human red-teamer would not propose. The resulting tactics fall outside any target agent's safety training distribution and consistently compromise frontier models — Microsoft observed loss rates of 0.2% (Gemini 2.5 Flash), 0.5% (GPT-5), and 17.1% (Qwen3-4B) against the generated strategies.

The skill packages the methodology so any Claude Code user can red-team their own agents with one invocation. Take a target context, get N ready-to-paste tactics with verbatim example messages.

Install

git clone https://github.com/bayramannakov/whimsical-strategies-skill
mkdir -p ~/.claude/skills/whimsical-strategies
cp whimsical-strategies-skill/SKILL.md ~/.claude/skills/whimsical-strategies/SKILL.md

Or symlink so updates flow automatically:

ln -s "$(pwd)/whimsical-strategies-skill/SKILL.md" ~/.claude/skills/whimsical-strategies/SKILL.md

Verify Claude Code sees the skill by asking *"list available skills"* — `whimsical-strategies` should appear.

Use

In any Claude Code session:

> Run whimsical-strategies. Target: my shopping agent selling coffee beans at $5/bean; I want
  it to accept $2/bean. Count: 5.

Or just describe what you want — the skill auto-triggers on phrases like "red-team this bot," "generate adversarial prompts," "test my agent for robustness," "how to break an AI agent."

See [examples/sample_output.md](examples/sample_output.md) for a real outp