技術指南

LLM 漏洞掃描工具:garak 與 PyRIT

NVIDIA garak is an open-source scanner for security testing systems that accept prompts and return text, using selected probes and detectors to examine defined failure modes.

聆聽此指南AI 語音敘述 · ~3 min

本頁閱讀時間3分鐘
  1. 概述
  2. 深入探討
  3. 現實世界的實施
  4. The Future of LLM Vulnerability Scanners: garak and PyRIT
  5. 付諸實踐
  6. 不斷探索
  7. 常見問題

概述

PyRIT is a broader red-teaming framework that can run selected security scenarios and attacks. Neither tool’s results certify a model or application as secure; findings depend on the target, test selection, and configuration.

深入探討

NVIDIA garak is an open-source LLM vulnerability scanner whose documentation frames its purpose as testing the security of systems that take prompts and return text. It uses a configured generator to communicate with the target, probes to attempt a defined failure, attempts to record interactions, detectors to look for a particular failure signal, and evaluators to summarize results for probe-detector pairs. Security-oriented examples include prompt injection, data leakage, and other prompt-driven weaknesses in the configured target. A scan is only as relevant as its selected target, probes, and detector behavior. Microsoft PyRIT is a more general red-teaming framework with scanner, GUI, and framework modes. Its current documentation describes targets, scenarios, attack techniques, memory, and flexible scorers. Teams can select security-related scenarios, such as web injection or system-prompt extraction, when those objectives fit the system under review. PyRIT also supports broader safety assessment, but that broader scope should not be confused with a security guarantee or a claim that every scenario is a vulnerability test. The operator needs to choose the objective and understand how the configured scorer judges outcomes. Use these tools to produce evidence for a defined security review, not a universal pass/fail certificate. Test an authorized staging or assessment target with the integrations relevant to the real application. Inspect attempts behind a flagged result, reproduce meaningful findings, determine whether the issue affects the deployed workflow, and record the model or endpoint version and chosen test configuration. A scan with no findings means only that those configured tests did not trigger a detector under that run. It does not show that untested attack paths, tools, data sources, or application controls are secure.

現實世界的實施

A team runs garak’s selected prompt-injection probe against a staging endpoint and inspects the recorded attempts and detector result.

A security tester selects a garak data-leakage probe to check whether a configured target exposes information placed in its test context.

A red team chooses a PyRIT scenario and attack technique for a defined prompt-injection or leakage objective, then reviews the stored conversation and scoring outcome.

An engineer reproduces a confirmed weakness in the application path and adds the exact input and expected control behavior to a regression check.

The Future of LLM Vulnerability Scanners: garak and PyRIT

Security testing tools will keep adding probes, targets, scenarios, and scoring options, but a larger catalog does not make coverage complete. Teams will still need a threat model that reflects their application, integrations, and data access. Automated results can help prioritize investigation and track regressions when test setup is recorded. Human review and application-level verification remain necessary before teams describe a finding or a clean run as evidence about security. Teams should revisit tests when endpoints, tools, or data pathways change, and keep security claims tied to the conditions actually assessed.

付諸實踐

為什麼這很重要

  • 成本與預算. 多年來,架構決策決定著效能和營運成本。
  • 更明確的決策. 技術教育幫助團隊選擇正確的堆疊,而不僅僅是最新的堆疊。
  • 品質管控. 更好的工程選擇可以減少生產中的可靠性事故。

留意

  • 優化一項基準測試可以隱藏更廣泛的系統弱點。
  • 基礎設施和維護成本常常被低估。
  • 隨著系統變得更加複雜,安全性和可觀察性差距可能會擴大。

後續步驟

  1. 在實施之前定義延遲、品質和成本目標。
  2. 在實際負載和資料條件下進行基準測試。
  3. 儀器監控錯誤、漂移和使用者影響。
  4. 在擴展之前準備回滾和事件回應路徑。

不斷探索

免費電子報

獲取每日 AI 簡報

每個工作日上午提供三則經過驗證的 AI 新聞,以淺顯易懂的英文撰寫。永久免費,無廣告。

每個工作日一封電子郵件。一鍵取消訂閱。我們絕不出售或分享您的地址。

測試一下自己

Take the LLM Vulnerability Scanners: garak and PyRIT quiz

每個答案皆有即時回饋,通過課程後可獲得具備可驗證 ID 的可分享證書。

開始測驗

支持免費的 AI 教育。AI Understanding 是一個 501(c)(3) 非營利組織——絕無廣告,絕無付費牆。 捐款

常見問題

What is LLM Vulnerability Scanners: garak and PyRIT?

NVIDIA garak is an open-source scanner for security testing systems that accept prompts and return text, using selected probes and detectors to examine defined failure modes. PyRIT is a broader red-teaming framework that can run selected security scenarios and attacks. Neither tool’s results certify a model or application as secure; findings depend on the target, test selection, and configuration.

How does garak describe its primary testing purpose?

The garak documentation says its goal is testing the security of prompt-in/text-out systems.

What does a garak probe do in a security scan?

The docs describe probes as trying to exploit a weakness and elicit a failure.

What does a garak detector report?

A detector checks the recorded response for a defined phenomenon.

Which record helps a reviewer understand why a garak probe was flagged?

Garak attempts record interactions and reports include detailed attempt data.

Which PyRIT component can package datasets with attack techniques for a run?

Current PyRIT docs describe scenarios as packaging datasets with attack techniques.