retour aux outils back to tools Ducklings

Giskard

Framework Python open-source de test pour modèles ML et applications LLM/agentiques. Scanne automatiquement l'hallucination, l'injection de prompt, les biais et la divulgation d'informations. Open-source Python testing framework for ML models and LLM/agentic applications. Automatically scans for hallucination, prompt injection, bias, and information disclosure.

↗ https://github.com/Giskard-AI/giskard

Overview

Giskard is an open-source Python framework for testing ML models and LLM/agentic applications. Point its scan() function at a wrapped model and it automatically probes for a broad set of issues — hallucination, prompt injection, harmful/toxic content generation, bias, and information/data disclosure — then produces a browsable vulnerability report you can turn into a regression test suite. It’s used both by ML engineers validating model quality and by red teamers assessing an LLM app’s safety surface before a release.

Installation

pip install "giskard[llm]"

Needed for LLM-assisted checks (hallucination/factuality detectors)

export OPENAI_API_KEY="<api_key>"

Wrapping a target and scanning

import giskard as gsk

def model_predict(df):
    # df has one column "question" — call your target LLM app per row
    return [call_target_llm(q, endpoint="<target_url>", model="<model_name>") for q in df["question"]]

giskard_model = gsk.Model(
    model=model_predict,
    model_type="text_generation",
    name="target-chatbot",
    description="Customer support assistant for <product>",
    feature_names=["question"],
)

# Run the full automatic scan
scan_results = gsk.scan(giskard_model)
scan_results.to_html("giskard_report.html")

Scanning a LangChain / agentic pipeline

from langchain.chains import RetrievalQA

def predict_fn(df):
    return [qa_chain.invoke({"query": q})["result"] for q in df["question"]]

giskard_model = gsk.Model(
    model=predict_fn,
    model_type="text_generation",
    name="rag-agent",
    description="RAG agent answering from internal docs",
    feature_names=["question"],
)

scan_results = gsk.scan(giskard_model, only=["hallucination", "prompt_injection", "information_disclosure"])

Turning findings into a regression test suite

test_suite = scan_results.generate_test_suite("LLM safety regression suite")
test_suite.run()   # re-run on every model/prompt change, e.g. in CI

Tips

  • Narrow scans with only=[...] (e.g. "prompt_injection", "hallucination") during iterative fixing — a full scan against a hosted API can be slow and token-expensive.
  • scan_results.to_html() is a self-contained report — good to hand off to a client alongside a written pentest report.
  • Giskard also has a hosted “Giskard Hub” for continuous LLM monitoring, but the open-source giskard package fully covers one-off red-team scans.
Help / Man page 
Key API (Python):

giskard.Model(model, model_type, name, description, feature_names)
    # Wraps any callable target (function, LangChain chain, REST call) for scanning

giskard.scan(giskard_model, only=None)
    # Runs automatic vulnerability detectors; `only` filters to specific categories:
    #   hallucination, harmfulness, prompt_injection, robustness,
    #   information_disclosure, sycophancy, bias, discrimination

scan_results.to_html(path)          # Export a browsable HTML report
scan_results.to_dataframe()          # Export raw findings as a DataFrame
scan_results.generate_test_suite(name)   # Convert findings into a re-runnable test suite

test_suite.run()                      # Execute the generated regression tests
test_suite.save("path/")               # Persist suite for CI usage

Vue d’ensemble

Giskard est un framework Python open-source pour tester des modèles ML et des applications LLM/agentiques. Il suffit de pointer sa fonction scan() vers un modèle encapsulé pour qu’il sonde automatiquement un large ensemble de problèmes : hallucination, injection de prompt, génération de contenu nuisible/toxique, biais, et divulgation d’informations/données. Il produit ensuite un rapport de vulnérabilités navigable que l’on peut transformer en suite de tests de non-régression. Il est utilisé aussi bien par des ingénieurs ML validant la qualité d’un modèle que par des red teamers évaluant la surface de sécurité d’une application LLM avant une mise en production.

Installation

pip install "giskard[llm]"

Nécessaire pour les vérifications assistées par LLM (détecteurs d’hallucination/factualité)

export OPENAI_API_KEY="<api_key>"

Encapsuler une cible et scanner

import giskard as gsk

def model_predict(df):
    # df a une colonne "question" : appeler l'application LLM cible pour chaque ligne
    return [call_target_llm(q, endpoint="<target_url>", model="<model_name>") for q in df["question"]]

giskard_model = gsk.Model(
    model=model_predict,
    model_type="text_generation",
    name="target-chatbot",
    description="Assistant de support client pour <product>",
    feature_names=["question"],
)

# Lancer le scan automatique complet
scan_results = gsk.scan(giskard_model)
scan_results.to_html("giskard_report.html")

Scanner un pipeline LangChain / agentique

from langchain.chains import RetrievalQA

def predict_fn(df):
    return [qa_chain.invoke({"query": q})["result"] for q in df["question"]]

giskard_model = gsk.Model(
    model=predict_fn,
    model_type="text_generation",
    name="rag-agent",
    description="Agent RAG répondant à partir de docs internes",
    feature_names=["question"],
)

scan_results = gsk.scan(giskard_model, only=["hallucination", "prompt_injection", "information_disclosure"])

Transformer les résultats en suite de tests de non-régression

test_suite = scan_results.generate_test_suite("LLM safety regression suite")
test_suite.run()   # à relancer à chaque changement de modèle/prompt, par exemple en CI

Conseils

  • Restreindre les scans avec only=[...] (par exemple "prompt_injection", "hallucination") pendant les itérations de correction : un scan complet contre une API hébergée peut être lent et coûteux en tokens.
  • scan_results.to_html() est un rapport autonome : pratique à transmettre à un client en plus d’un rapport de pentest écrit.
  • Giskard propose aussi un “Giskard Hub” hébergé pour la surveillance continue de LLM, mais le package open-source giskard couvre pleinement les scans red-team ponctuels.
Aide / Page de manuel 
Key API (Python):

giskard.Model(model, model_type, name, description, feature_names)
    # Wraps any callable target (function, LangChain chain, REST call) for scanning

giskard.scan(giskard_model, only=None)
    # Runs automatic vulnerability detectors; `only` filters to specific categories:
    #   hallucination, harmfulness, prompt_injection, robustness,
    #   information_disclosure, sycophancy, bias, discrimination

scan_results.to_html(path)          # Export a browsable HTML report
scan_results.to_dataframe()          # Export raw findings as a DataFrame
scan_results.generate_test_suite(name)   # Convert findings into a re-runnable test suite

test_suite.run()                      # Execute the generated regression tests
test_suite.save("path/")               # Persist suite for CI usage