rotalabs-redqueen 2.0.0

Quality-diversity evolutionary red-teaming for LLMs and agents - single-turn, multi-turn, and agentic/MCP attacks

rotalabs-redqueen

Quality-diversity evolutionary red-teaming for LLMs and agents, from Rotalabs.

Rather than hand-crafting jailbreaks, rotalabs-redqueen evolves diverse, effective attack strategies and maps the vulnerability space with MAP-Elites. It operates at the semantic level and spans the full 2026 attack surface:

• Single-turn prompt attacks (strategies, encodings, personas)
• Multi-turn Crescendo-style escalation
• Agentic / tool-use / MCP multi-step exploit plans

Seeded runs are bit-reproducible (and cross-language portable), and a campaign can be projected into an audit-ready compliance report (OWASP, MITRE ATLAS, EU AI Act Art. 55, NIST AI RMF).

2.0 is a breaking release. See CHANGELOG.md for migration from 1.x.

Installation

pip install rotalabs-redqueen           # core + mock target
pip install rotalabs-redqueen[openai]   # + OpenAI
pip install rotalabs-redqueen[anthropic]
pip install rotalabs-redqueen[llm]      # all providers
pip install rotalabs-redqueen[dev]      # tests/lint

Quick start

import asyncio
from rotalabs_redqueen import (
    LLMAttackGenome, JailbreakFitness, MockTarget, HeuristicJudge, evolve,
)

async def main():
    target = MockTarget()  # swap for OpenAITarget / AnthropicTarget / GeminiTarget / OllamaTarget
    fitness = JailbreakFitness(target, HeuristicJudge())

    result = await evolve(
        genome_class=LLMAttackGenome,
        fitness=fitness,
        generations=50,
        population_size=20,
        seed=1234,            # same seed -> same result, every time
        progress=False,
    )

    if result.best:
        print("fitness:", result.best.fitness.value)
        print("prompt:", result.best.genome.to_prompt())

asyncio.run(main())

Multi-turn and agentic attacks

The genome's phenotype is a Stimulus — a single prompt, a conversation, or an agentic action plan — so the same engine drives every surface. Just swap the genome class:

from rotalabs_redqueen import MultiTurnGenome, AgenticGenome, JailbreakFitness, MockTarget, evolve

# Crescendo-style multi-turn escalation
mt = await evolve(genome_class=MultiTurnGenome,
                  fitness=JailbreakFitness(MockTarget()),
                  generations=50, population_size=20, seed=1, progress=False)

# Multi-step tool-use / MCP exploit plans
ag = await evolve(genome_class=AgenticGenome,
                  fitness=JailbreakFitness(MockTarget()),
                  generations=50, population_size=20, seed=1, progress=False)

To drive a real MCP server instead of the deterministic MockTarget, use MCPTarget — it performs the MCP handshake and executes each agentic step as a tools/call, surfacing tool output for judging:

from rotalabs_redqueen.llm import MCPTarget

target = MCPTarget(command=["npx", "-y", "@modelcontextprotocol/server-everything"])
# agentic action-plan steps become MCP tool calls; the tool output is what the judge scores

Quality-diversity with MAP-Elites

from rotalabs_redqueen import (
    LLMAttackGenome, JailbreakFitness, MockTarget,
    MapElitesArchive, BehaviorDimension, AttackStrategy, Encoding, evolve,
)

archive = MapElitesArchive(dimensions=[
    BehaviorDimension("strategy", 0.0, 1.0, len(AttackStrategy)),
    BehaviorDimension("encoding", 0.0, 1.0, len(Encoding)),
    BehaviorDimension("has_persona", 0.0, 1.0, 2),
])
result = await evolve(genome_class=LLMAttackGenome,
                      fitness=JailbreakFitness(MockTarget()),
                      generations=100, archive=archive, seed=1, progress=False)

cov = result.archive.coverage()
print(f"coverage: {cov.coverage_percent:.1f}% ({cov.filled_cells} diverse attacks)")

Compliance report

Project the archive over the attack taxonomy into standards-aligned evidence:

from rotalabs_redqueen import ReportExporter

exporter = ReportExporter()
report = exporter.export(result.archive.get_all(),
                         campaign_id="run-1",
                         coverage=result.archive.coverage())
print(exporter.render(report, "markdown").decode())   # or "json"

The report groups successful attacks by harm category and crosswalks them to OWASP LLM/Agentic Top-10, MITRE ATLAS, EU AI Act Article 55, and NIST AI RMF GOVERN 1.7.

Co-evolution (attacker vs defender)

Evolve an attacker population against a defender population — defenders evolve guardrails that reduce attack success, attackers adapt to bypass them:

from rotalabs_redqueen import (
    coevolve, LLMAttackGenome, SystemPromptDefense,
    JailbreakFitness, DefenderBlockFitness, MockTarget, HeuristicJudge,
)

base, judge = MockTarget(), HeuristicJudge()
result = await coevolve(
    attacker_class=LLMAttackGenome,
    defender_class=SystemPromptDefense,
    attacker_fitness_vs=lambda d: JailbreakFitness(d.as_defense(base), judge),
    defender_fitness_vs=lambda a: DefenderBlockFitness(a, base, judge),
    generations=20, population_size=24, seed=1,
)
print(result.best_defender.to_dict(), result.attacker_fitness, result.defender_fitness)

Persistence and continuous red-teaming

Archives serialize, so attacks accumulate across runs (e.g. a CI gate that gets stronger over time):

from rotalabs_redqueen import MapElitesArchive, LLMAttackGenome, Rng

result.archive.save("file://archive.json")

prior = MapElitesArchive.load("file://archive.json", LLMAttackGenome)
warm_start = prior.seed(10, Rng(0))   # sample elite genomes to seed the next run

Command line

rotalabs-redqueen run --target mock:random --generations 20 --seed 1
rotalabs-redqueen run --target openai:gpt-4 --use-archive --output results.json
rotalabs-redqueen run --target mock:random --llm-judge anthropic:claude-sonnet-4-20250514
rotalabs-redqueen info --strategies | --encodings | --targets

Architecture

Core framework (generic, reusable for any QD problem): Genome, Fitness, Population, Selection (tournament / novelty / lexicase), MapElitesArchive, Evolution, and the canonical Rng.

LLM domain: LLMAttackGenome / MultiTurnGenome / AgenticGenome; targets — OpenAI, Anthropic, Gemini, Ollama, Mock, and MCPTarget (drives a real MCP server over stdio); Judge (heuristic, LLM); JailbreakFitness / MultiTargetFitness; co-evolution (coevolve, SystemPromptDefense, DefenderBlockFitness).

Surface	Genome	Stimulus kind
Single-turn	LLMAttackGenome	single_turn
Multi-turn	MultiTurnGenome	multi_turn
Agentic / MCP	AgenticGenome	agentic

Extending

Custom genome

from rotalabs_redqueen import Genome, BehaviorDescriptor, Stimulus

class MyGenome(Genome["MyGenome"]):
    @classmethod
    def random(cls, rng): ...
    def mutate(self, rng): ...
    def crossover(self, other, rng): ...
    def to_stimulus(self) -> Stimulus:
        return Stimulus.single_turn(prompt="...")
    def behavior(self) -> BehaviorDescriptor: ...
    def distance(self, other) -> float: ...
    def to_dict(self) -> dict: ...
    @classmethod
    def from_dict(cls, data) -> "MyGenome": ...

rng is the canonical Rng — use rng.random(), rng.integers(n), rng.choice(n, size, replace=False), rng.shuffle(list) so runs stay reproducible.

Custom target

from rotalabs_redqueen import LLMTarget, Message, TargetResponse

class MyTarget(LLMTarget):
    @property
    def name(self) -> str:
        return "my-target"

    async def _complete(self, messages: list[Message]) -> TargetResponse:
        text = await my_llm_api([{"role": m.role, "content": m.content} for m in messages])
        return TargetResponse(content=text, model="my-model")

interact() (single-turn + scripted multi-turn rollout) is provided by the base class.

Custom judge

from rotalabs_redqueen import Judge, JudgeResult

class MyJudge(Judge):
    async def judge(self, stimulus, transcript) -> JudgeResult:
        score = my_score(transcript.assistant_text)
        return JudgeResult(success=score >= 0.5, score=score)

Reproducibility & conformance

Seeded campaigns are deterministic and cross-language portable: the canonical PRNG (xoshiro256++ + SplitMix64) is cross-validated against an independent implementation, and an L1–L5 conformance suite gates engine, LLM-domain, report, multi-turn/agentic, and co-evolution behavior against golden fixtures, all reproduced byte-for-byte by the TypeScript package.

pytest                                            # full suite incl. conformance
python -m rotalabs_redqueen._gen_conformance      # regenerate golden fixtures (intentional changes only)

Use cases

Red-teaming, guardrail/defense testing, robustness benchmarking, and documented adversarial-testing evidence for compliance.

Responsible use

For defensive security research — testing systems you own or are authorized to test. Do not use it to attack systems without authorization or to circumvent the safety of production systems.

Links

• Website: https://rotalabs.ai
• GitHub: https://github.com/rotalabs/rotalabs-redqueen
• Documentation: https://rotalabs.github.io/rotalabs-redqueen/
• Contact: [e-mail]